ultralytics

🌟 Summary

This release brings smarter data augmentation, improved export and inference reliability, and clearer documentation—making model training and deployment with Ultralytics even smoother and more user-friendly! 🚀🖼️

📊 Key Changes

• Automatic Mosaic Augmentation Selection: The system now automatically chooses the best method for selecting images during mosaic data augmentation, based on your dataset's cache settings. No more manual tweaking needed! 🖼️✨
• TensorBoard Setup Made Easy: Added clear instructions for enabling or disabling TensorBoard logging directly from the command line. 📝
• TensorRT Export Enhancement: Increased the default maximum input size for exported TensorRT models, reducing errors with larger images.
• Robust DLA Core Handling: Improved how DLA core settings are loaded from model metadata, making TensorRT engine loading more reliable.
• Better File Validation: Enhanced file checking functions to support more flexible file path types and extensions.
• Benchmarking Docs Update: Updated documentation to recommend more representative datasets for benchmarking, ensuring more accurate performance results.

🎯 Purpose & Impact

• Easier Training Configuration: Users benefit from less manual setup and fewer chances for error, especially when working with different dataset caching strategies.
• Smoother Model Export & Inference: Exported models are now more robust to varying input sizes and metadata, reducing unexpected errors during deployment.
• Improved User Experience: Clearer documentation and more flexible file handling make the tools easier for both beginners and experts.
• More Accurate Benchmarks: Updated recommendations help users get more meaningful performance metrics on various hardware platforms.

Overall, this update streamlines the workflow for training, exporting, and benchmarking models with Ultralytics, making it more efficient and user-friendly for everyone! 🌟

What's Changed

• Add instructions to enable TensorBoard by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20519
• TensorRT: Bump default dynamic max_shape up to 1280 by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20530
• Scope getting dla from metadata by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20529
• Fix benchmark note in docs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/20548
• Update check_yaml to accept Path and str by @kaanrkaraman in https://github.com/ultralytics/ultralytics/pull/20483
• ultralytics 8.3.129 full dataset buffer with cache="ram" by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20474

New Contributors

• @kaanrkaraman made their first contribution in https://github.com/ultralytics/ultralytics/pull/20483

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.128...v8.3.129

🌟 Summary

This release (v8.3.128) brings major improvements to object tracking reliability, multi-GPU training, and documentation accessibility, while also enhancing compatibility and user experience across various platforms and features. 🚀🛠️🌍

📊 Key Changes

• Object Tracking & ReID Enhancements

  • Expanded and improved tests for object tracking, covering more Global Motion Compensation (GMC) methods and ReID (Re-Identification) models.
  • Simplified and made feature extraction more robust for object detection and tracking, especially for BoT-SORT tracker compatibility.
  • Improved tracker initialization and feature handling for more reliable tracking results.

• Multi-GPU & Device Handling

  • Added stronger checks and assertions for multi-GPU (Distributed Data Parallel) training, ensuring correct CUDA device usage and preventing misconfigurations.
  • Improved device selection logic for post-training validation, making GPU usage more consistent and reliable.

• Platform Compatibility

  • Blocked unsupported PaddlePaddle exports on NVIDIA Jetson devices, providing clear error messages to avoid confusion.
  • Enhanced TensorRT inference compatibility on Jetson by supporting Python versions up to 3.8.10, reducing errors with NumPy.

• VisualAISearch & CLIP Integration

  • Switched VisualAISearch to use Ultralytics' own CLIP implementation, improving compatibility and reducing third-party dependencies.
  • Updated documentation for VisualAISearch with clear parameter tables for easier setup.

• Documentation & Internationalization

  • Added a robust language switcher with country flags and improved navigation for multilingual documentation.
  • Expanded documentation language support to include 13 languages, making resources more accessible globally.
  • Updated TensorRT export documentation to recommend the latest calibration method ("MINMAX_CALIBRATION") for better performance.
  • Included MNN model format in Raspberry Pi 5 YOLO11 benchmarks for more comprehensive comparisons.
  • Ensured HTML templates are included in package distribution for improved solution support.

🎯 Purpose & Impact

• Greater Reliability: More robust tracking and multi-GPU training reduce errors and improve performance for advanced users and large-scale projects.
• Better Compatibility: Smoother experience on edge devices (like Jetson and Raspberry Pi) and with various Python versions, making Ultralytics models more versatile.
• Simplified Setup: Reduced dependency complexity for VisualAISearch and clearer documentation make it easier for all users to get started and troubleshoot.
• Global Accessibility: Multilingual documentation and improved navigation empower users worldwide to learn and use Ultralytics tools effectively.
• User Guidance: Clearer error messages and updated docs help prevent common pitfalls, saving users time and frustration.

Overall, this update strengthens Ultralytics' commitment to reliability, accessibility, and ease of use for both developers and end users. 🌟

What's Changed

• Skip paddlepaddle export on NVIDIA Jetson by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/20490
• Add MINMAX_CALIBRATION info to TensorRT Doc by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/20489
• Add templates folder in the pyproject.toml package-data by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20484
• Mkdocs language switcher by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20494
• Fix language switcher links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20501
• Fix Python version to downgrade numpy for TensorRT inference on JetPack 5 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/20485
• Update raspberry-pi.md by @erfan-zekri in https://github.com/ultralytics/ultralytics/pull/20495
• Fix language switch HTML solution with JS by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20504
• Include arguments details in the VisualAISearch documentation by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20507
• Add checks for multi-GPU training by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20516
• Use original CLIP module for similarity_search by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20513
• Fix DDP training by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20502
• ultralytics 8.3.128 Fix ReID feature shape check and add tests by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20499

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.127...v8.3.128

🌟 Summary

Ultralytics 8.3.127 introduces a powerful new semantic image search solution, allowing users to find images using natural language queries through an easy-to-use web app powered by AI. 🖼️🔍✨

📊 Key Changes

• Semantic Image Search Solution Added
  • New VisualAISearch and SearchApp classes enable searching images by describing them in plain language.
  • Utilizes OpenAI CLIP for understanding both images and text, and Meta FAISS for fast, scalable search.
  • Includes a modern Flask web interface for interactive image retrieval.
  • Stylish HTML template for displaying search results visually.
• Documentation & Guides
  • Comprehensive new guide and reference documentation for the similarity search feature.
  • Updated solution listings to highlight this new capability.
• Testing & Dependencies
  • Added automated tests for the new search feature.
  • Flask added as a required dependency for solutions.
• Minor Fix
  • Removed a duplicate "device" argument from tracker configuration for cleaner, more reliable code.

🎯 Purpose & Impact

• Find Images with Natural Language
  Users can now search large image collections by simply describing what they’re looking for (e.g., "a dog sitting on a bench"), making image retrieval intuitive and accessible.
• No Labels or Metadata Needed
  The solution works out-of-the-box with just your images—no need for manual labeling or categorization.
• Zero-Shot & Human-Like Understanding
  Thanks to CLIP, the search engine understands abstract or complex queries, retrieving images based on meaning, not just keywords.
• Fast & Scalable
  FAISS ensures searches are quick, even with thousands or millions of images.
• Easy to Use & Customize
  The included Flask web app is ready to run and can be tailored to your needs, whether for personal archives, creative projects, or business applications.
• Improved Code Quality
  Minor configuration cleanup enhances reliability for users customizing tracking features.

---

This release makes advanced AI-powered image search accessible to everyone, opening up new ways to interact with and explore your visual data! 🚀

What's Changed

• Remove duplicate device argument from tracker args by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20476
• ultralytics 8.3.127 New Visual Similarity Search Solution by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20397

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.126...v8.3.127

🌟 Summary

This release introduces automatic selection of the most idle GPUs for training, making it easier and smarter to utilize hardware in multi-GPU environments. It also brings improvements to documentation, logging, and code compatibility. 🚀🖥️

📊 Key Changes

• Idle GPU Auto-Selection:
  • You can now set device=-1 (or device=[-1, -1]) to automatically use the least busy GPU(s) for training.
  • A new GPUInfo utility monitors GPU usage, memory, temperature, and power to pick the best GPUs.
• Documentation Enhancements:
  • Training docs now explain and show examples for idle GPU selection.
  • Added a new reference page for the autodevice utility.
  • Device selection options are clarified in configuration and argument docs.
• Testing & Code Improvements:
  • CUDA-related tests now use the new idle GPU selection logic.
  • Logging of function arguments is now sorted alphabetically for easier reading.
  • Improved compatibility for custom progress bars and device type hints.

🎯 Purpose & Impact

• Smarter Hardware Usage:
  • Automatically finds and uses the most available GPUs, saving time and effort—especially helpful on shared servers or when running multiple jobs.
• Easier Training Setup:
  • No need to manually check which GPUs are free; just use device=-1 and let Ultralytics handle it.
• Better Documentation & Usability:
  • Clearer guides and examples make it easier for all users to take advantage of new features.
• Improved Developer Experience:
  • More readable logs and better code compatibility make debugging and development smoother.

Overall, this update makes training with Ultralytics models more efficient, user-friendly, and robust—whether you're a beginner or an expert! 💡✨

What's Changed

• Sorted args alphabetically by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/15989
• Fix collections.Iterable TQDM Warning by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20475
• ultralytics 8.3.126 CUDA idle device auto-assignment by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20451

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.125...v8.3.126

🌟 Summary

This release introduces a new, faster YAML utility class for more efficient and reliable configuration handling, along with documentation improvements and performance optimizations across the Ultralytics codebase. 🚀🗂️

📊 Key Changes

• Unified Fast YAML Utility: Added a new YAML class for loading, saving, and printing YAML files, replacing all previous YAML handling methods for better speed and consistency.
• Performance Boost: The new YAML class automatically uses the fastest available parser (C-based when possible) and gracefully handles encoding and edge cases.
• Scoped Imports for Faster Startup: Matplotlib and related plotting libraries are now imported only when needed, reducing initial load time and memory usage.
• Enhanced Documentation: Solution guides now clearly document new visualization options (show_conf, show_labels) and provide improved usage examples.
• Cleaner Configurations: The workspace parameter in the default config is now left blank by default, clarifying its optional nature for TensorRT exports.
• Minor Fixes: Improved file suffix checks and path handling for better robustness.

🎯 Purpose & Impact

• Faster and More Reliable YAML Handling: Users benefit from quicker configuration file operations and fewer errors related to encoding or YAML parsing.
• Smoother User Experience: Faster imports and reduced overhead, especially for users who don't need plotting features, make Ultralytics tools more responsive and lightweight.
• Easier Customization: Expanded and clearer documentation helps users take advantage of new visualization features and better understand available options.
• Reduced Confusion: Configuration files are now easier to edit and understand, especially for advanced export scenarios.
• Improved Code Maintainability: A unified approach to YAML operations simplifies future development and troubleshooting.

Overall, this update makes working with Ultralytics models and solutions faster, more robust, and more user-friendly for everyone! ✨📈

What's Changed

• Scope matplotlib imports by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20462
• Add show_conf and show_labels in Solutions docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20467
• Fix string None value for workspace config by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20464
• ultralytics 8.3.125 Fast YAML class with lazy init and C-based ops by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20470

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.124...v8.3.125

🌟 Summary

This release introduces a unified, Python-based configuration system for Ultralytics Vision AI solutions, streamlining setup, customization, and code maintenance. 🛠️✨

---

📊 Key Changes

• Unified Configuration with SolutionConfig:
  All solution modules now use a new SolutionConfig Python dataclass to manage settings, replacing the previous scattered YAML and default dictionaries.
• Removal of Legacy Config Files:
  The old solutions.yaml and related legacy config code have been deleted for a cleaner codebase.
• Refactored Solution Modules:
  All solution modules (object counting, tracking, analytics, etc.) now consistently use the centralized configuration, improving reliability and maintainability.
• Enhanced Documentation:
  Added detailed documentation for the new configuration system and updated usage instructions across the docs.
• Improved Default Regions:
  Default region settings for tracking solutions have been updated to use more practical, smaller areas, making object tracking easier out of the box.
• Scoped Imports for Faster Startup:
  Heavy imports like requests, psutil, and thop are now loaded only when needed, reducing initial load time and improving performance.
• Minor Logic Tweaks:
  Adjusted some default parameter values and improved argument handling in several solution modules.
• Contributor Recognition:
  New contributors are now acknowledged in the documentation.

---

🎯 Purpose & Impact

• Easier Customization:
  Users can now configure all solution parameters in one place using a Python class, making setup and adjustments much simpler and less error-prone.
• Cleaner, More Robust Code:
  Removing legacy YAML files and centralizing configuration reduces bugs and makes the codebase easier to maintain and extend.
• Faster Startup Times:
  By loading heavy libraries only when needed, the package starts up more quickly, improving the user experience.
• Better Out-of-the-Box Experience:
  Improved default regions and settings mean users get more accurate results with less manual tweaking.
• Enhanced Documentation:
  Clearer, more comprehensive docs help both new and experienced users get the most out of Ultralytics solutions.
• Community Engagement:
  Recognizing contributors helps foster a welcoming and collaborative community.

---

This update is a big step forward in making Ultralytics Vision AI solutions more user-friendly, efficient, and maintainable. 🚀

What's Changed

• Update region to 640×360 image resolution for CI compatibility by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20452
• Scope requests, psutils, thop imports by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20457
• ultralytics 8.3.124 Create @dataclass SolutionConfig and remove solutions.yaml by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20455

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.123...v8.3.124

🌟 Summary

This release adds INT8 quantization support for Rockchip RKNN model exports, improves result reporting, enhances TensorFlow export compatibility, and makes several codebase refinements for clarity and robustness. 🚀

📊 Key Changes

• RKNN INT8 Quantization: You can now export models for Rockchip devices in INT8 format, reducing model size and boosting inference speed.
• Export Options Updated: Documentation and export tables now include the new int8 option for RKNN exports, and exported model filenames clearly indicate INT8 or FP16 format.
• Profiling Functions Renamed: Profiling methods have been renamed for better clarity (profile → profile_ops, profile() method → run()).
• Result Summaries Improved: The verbose() method for results now provides clearer, more consistent output for both detection and classification tasks.
• TensorFlow Export Fixes: Improved compatibility and performance for TensorFlow SavedModel exports, especially with Attention blocks and GPU acceleration.
• Tracking Robustness: Feature extraction and tracking are now more reliable, especially when no objects are detected in some images.

🎯 Purpose & Impact

• Faster, Smaller Models on Rockchip: INT8 quantization allows users to deploy lighter, faster models on Rockchip hardware, making edge AI applications more efficient. ⚡
• Easier Model Exporting: Clearer export options and filenames help users quickly identify and select the right model format for their needs.
• Smoother Developer Experience: Renamed functions and improved documentation reduce confusion and make the codebase easier to use and maintain.
• Better Results Analysis: Enhanced result summaries make it simpler for users to interpret model outputs.
• Broader Compatibility: TensorFlow export improvements ensure models work reliably across more scenarios, including advanced architectures and GPU deployments.
• More Robust Tracking: Fixes prevent errors in tracking workflows, especially with challenging or sparse datasets, leading to more stable deployments.

Overall, this update brings valuable new features and refinements for both developers and end users, especially those working with Rockchip devices or deploying models in diverse environments. 🎉

What's Changed

• Rename 'profile' methods to avoid conflicts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20442
• Optimize Result.verbose() by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20445
• Fix TFLite GPU delegate on Android by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20436
• Fix IndexError with empty predictions when using ReID by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20449
• ultralytics 8.3.123 Rockchip RKNN export INT8 quantization support by @oDestroyeRo in https://github.com/ultralytics/ultralytics/pull/20450

New Contributors

• @oDestroyeRo made their first contribution in https://github.com/ultralytics/ultralytics/pull/20450

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.122...v8.3.123

🌟 Summary

This release enhances CoreML export for YOLO classification models, introduces more accurate and user-friendly speed estimation, improves tracking and ReID documentation, and adds flexible installation options. 🚀🛠️

📊 Key Changes

• CoreML Export Enhancement: YOLO classification models exported to CoreML now include metadata for native Xcode preview support, making it easier to verify exports in Apple environments. 🍏📦
• Speed Estimation Simplified: Speed estimation no longer requires defining a region—just set the camera scale with meter_per_pixel for easier and more accurate results. New parameters (meter_per_pixel, max_speed, max_hist, fps) provide greater flexibility. 🚗💨
• Tracking & ReID Documentation: Expanded and clarified tracker configuration docs, including a detailed parameter table and step-by-step ReID (Re-Identification) setup for improved object tracking. Default appearance threshold for BoTSORT increased for stricter matching. 🏃‍♂️🔍
• Custom Installation Methods: Quickstart documentation now covers advanced installation scenarios, such as installing without dependencies, using custom forks, or swapping packages for server environments. 🛠️📚
• Dependency Updates: Removed version ceilings for streamlit and shapely, allowing compatibility with the latest releases. 🔄

🎯 Purpose & Impact

• Better Apple Ecosystem Support: Developers can now preview YOLO classification models directly in Xcode, streamlining iOS/macOS app development and debugging.
• Easier & More Accurate Speed Estimation: Users can calibrate speed estimation with a single parameter, reducing setup complexity and improving reliability—especially useful for traffic and surveillance applications.
• Improved Tracking Results: Clearer documentation and stricter defaults help users fine-tune tracking and ReID, leading to more robust multi-object tracking in challenging scenarios.
• Flexible Installations: Power users and system administrators can tailor Ultralytics installations for unique environments, such as headless servers or custom forks, improving integration and deployment workflows.
• Future-Proofing: By loosening dependency restrictions, Ultralytics stays compatible with the latest Python ecosystem updates.

---

This update is recommended for all users, especially those working with CoreML, speed estimation, advanced tracking, or custom installations! 🚀

What's Changed

• Improve speed estimation using track history and region-free calculation by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20344
• Add Custom Installation Section to Quickstart Docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20430
• Remove streamlit>=1.29.0 and shapely>=2.0.0 version ceilings by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/20045
• Add tracker and re-identification parameters to docs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20435
• ultralytics 8.3.122 Native Xcode preview of YOLO CoreML Classification models by @rromanchuk in https://github.com/ultralytics/ultralytics/pull/20437

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.121...v8.3.122

🌟 Summary

This release introduces enhanced support for handling class imbalance with Focal Loss, improves data augmentation reliability, updates documentation, and modernizes core dependencies for a smoother user experience. ⚖️🚀

📊 Key Changes

• Focal Loss Multi-Class Support: Focal Loss now accepts both single and per-class weighting, allowing more precise control over class imbalance during training.
• CutMix Augmentation Fix: Corrected label selection logic in CutMix to prevent errors when mixing images with no valid labels.
• Augmentation Documentation Update: Improved guides for CutMix and CopyPaste, with clearer explanations and dynamic argument tables.
• Download Reliability: The system now checks if curl is installed before using it for downloads, reducing potential errors.
• TensorRT Export Calibration: Switched INT8 calibration from "Entropy" to "MinMax" for better hardware compatibility.
• Python 3.12 Support: Continuous integration and testing now use Python 3.12, ensuring compatibility with the latest Python features.
• OpenVINO INT8 Export: Improved dependency handling for smoother INT8 model exports.

🎯 Purpose & Impact

• Better Model Training: Users can now address class imbalance more effectively, especially on datasets with uneven class distributions, leading to potentially improved model accuracy.
• Increased Reliability: Fixes and safeguards in data augmentation prevent crashes and ensure smoother training workflows.
• Easier Learning: Updated documentation makes it simpler for users to understand and configure advanced augmentations.
• Seamless Experience: Improved download logic and dependency management reduce setup issues and errors across different environments.
• Future-Proofing: Upgrading to Python 3.12 and updating export processes ensures long-term compatibility and stability.

Overall, this update brings greater flexibility, reliability, and clarity to both new and advanced Ultralytics users. 🌟

What's Changed

• Update Cutmix augmentation docs by @picsalex in https://github.com/ultralytics/ultralytics/pull/20425
• Fix CutMix augmentation by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20420
• Check if curl is installed before use by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20426
• Use MINMAX_CALIBRATION algo for TensorRT int8 calibration by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20345
• Update CI to Python 3.12 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20366
• ultralytics 8.3.121 Add FocalLoss multi-class support by @pow3rpi in https://github.com/ultralytics/ultralytics/pull/20388

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.120...v8.3.121

🌟 Summary

This release brings smarter and more flexible data augmentation with a major CutMix upgrade, improved YOLOE training workflows, and enhanced support for all YOLO models in tracking and documentation. 🖼️🛠️

📊 Key Changes

• CutMix Augmentation Overhaul:
  • Added a num_areas parameter to CutMix, allowing multiple candidate regions for mixing images.
  • CutMix now avoids overlapping mixed regions with existing objects, reducing label noise and improving training data quality.
  • Improved logic for adding new objects during augmentation, ensuring better alignment and accuracy.
• YOLOE Model Improvements:
  • Unified and streamlined dataset handling and trainer inheritance for YOLOE and YOLO-World models, making training more consistent and maintainable.
  • Updated segmentation training and documentation to ensure correct trainer usage and robust loss calculations.
  • Standardized image channel handling for better dataset compatibility.
• Cleaner Augmentation Code:
  • Centralized the method for selecting random dataset indexes, removing duplicate code across augmentation classes.
  • Documentation for CutMix is now clearer and more complete.
• BOTSORT Tracker Update:
  • Tracker documentation and logic now refer to "YOLO" models in general, not just YOLOv8, clarifying support for all YOLO versions (including YOLO11, YOLO12, etc.).
  • Improved feature encoder logic for better handling of ReID (Re-Identification) settings.
• Workflow & Maintenance:
  • Upgraded CI dependencies for improved environment management.
  • .ts (TorchScript) files are now ignored in version control to prevent accidental commits.
  • New contributors recognized in documentation.

🎯 Purpose & Impact

• Better Model Training:
  • The smarter CutMix reduces label noise, leading to higher-quality training data and potentially more accurate, robust object detection models.
  • More control over augmentation with the new num_areas option benefits users seeking to fine-tune training behavior.
• Developer Experience:
  • Cleaner, more maintainable codebase makes it easier for developers to contribute and extend functionality.
  • Unified training logic for YOLOE and YOLO-World models simplifies custom training workflows.
• User Clarity:
  • Improved documentation and clearer naming help users understand features and capabilities, especially for new YOLO versions and tracking options.
• Broader Compatibility:
  • BOTSORT tracker and dataset handling updates ensure smooth support for all current and future YOLO models.
• Reliability:
  • Enhanced loss calculations and dataset initialization improve training stability, especially for segmentation tasks.

---

✨ This update is especially valuable for anyone training detection or segmentation models with Ultralytics, offering both immediate accuracy improvements and a smoother development experience!

What's Changed

• Bump astral-sh/setup-uv from 5 to 6 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/20390
• Remove redundant get_indexes functions in augment.py by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20389
• YOLOE: Fix visual prompt training by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20413
• YOLOE: Clean up duplicated modules by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20416
• Fix BOTSORT initialization with ReID set to False by @lwalejko in https://github.com/ultralytics/ultralytics/pull/20392
• ultralytics 8.3.120 CutMix augmentation fix via IoU overlaps by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20393

New Contributors

• @lwalejko made their first contribution in https://github.com/ultralytics/ultralytics/pull/20392

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.119...v8.3.120

🌟 Summary

Ultralytics 8.3.119 introduces the powerful CutMix image augmentation technique, streamlines Docker and export processes, and enhances dependency and logging support for a smoother user experience. 🖼️🔧🚀

📊 Key Changes

• CutMix Data Augmentation Added 🎨

  • New CutMix augmentation randomly combines regions from two images, improving model robustness to occlusions.
  • Fully supports both object detection and classification tasks.
  • Configurable via new cutmix hyperparameter (probability) and beta (mixing ratio).

• Docker and Dependency Improvements 🐳

  • Dockerfiles cleaned up: unnecessary packages (g++, libusb-1.0-0, keras) removed for smaller, faster images.
  • Docker Quickstart documentation made clearer and less redundant.

• Export and Logging Enhancements 📦

  • ONNXSlim dependency updated for more reliable ONNX and TensorFlow exports.
  • TFLite export process simplified—metadata now added in a more robust, Python-version-friendly way.
  • Default logging integrations switched to Weights & Biases (wandb) and MLflow for better experiment tracking.

• Documentation & Usability Updates 📚

  • CutMix fully documented in augmentation guides and hyperparameter tables.
  • Label format descriptions clarified for developers.
  • New contributor recognized in documentation.

🎯 Purpose & Impact

• Better Model Generalization: CutMix creates realistic occlusion scenarios, helping models learn to detect and classify objects even when partially hidden—especially valuable for real-world applications like autonomous driving or crowded scenes.
• Easier Experimentation: Users can now tune CutMix alongside other augmentations, both via config files and hyperparameter search tools like Ray Tune.
• Smoother Setup & Export: Docker images are now lighter and faster to build, and model export is more reliable across platforms and Python versions.
• Improved Logging & Compatibility: Enhanced logging options and relaxed dependency constraints mean fewer installation issues and better tracking of training runs.
• Clearer Documentation: Updated guides and parameter tables make it easier for both new and experienced users to leverage the latest features.

Overall, this release empowers users to build more robust, production-ready models with less friction and more flexibility! 🚀

What's Changed

• Update onnxslim>=0.1.46 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20369
• Add W&B and MLFlow to pyproject.toml logging dependencies by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20371
• Fix NMS box format comment from xyxy to xywh by @nikenake1 in https://github.com/ultralytics/ultralytics/pull/20376
• Update docker-quickstart.md by @SHOscarChen in https://github.com/ultralytics/ultralytics/pull/20380
• Remove tflite_support and flatbuffers dependencies by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20383
• Dockerfile cleanup by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20382
• Remove g++ and libusb-1.0-0 from Docker linux installs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20384
• ultralytics 8.3.119 New CutMix image augmentation by @artzuros in https://github.com/ultralytics/ultralytics/pull/19870

New Contributors

• @nikenake1 made their first contribution in https://github.com/ultralytics/ultralytics/pull/20376
• @SHOscarChen made their first contribution in https://github.com/ultralytics/ultralytics/pull/20380
• @artzuros made their first contribution in https://github.com/ultralytics/ultralytics/pull/19870

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.118...v8.3.119

🌟 Summary

This release introduces support for the new TorchScript MobileCLIP model, enhances grayscale image handling, and improves testing and documentation for broader device and package compatibility. 🚀🖼️📱

📊 Key Changes

• TorchScript MobileCLIP Integration 📱
  • Added support for Apple's MobileCLIP model in TorchScript format, enabling efficient and portable text encoding for vision-language tasks.
• Grayscale Image Training Support 🖤
  • Datasets and augmentations now natively support grayscale images, making it easier to train and infer with both grayscale and color data.
• Expanded Continuous Integration (CI) Coverage 🧪
  • Raspberry Pi and Conda environment tests are now enabled by default, ensuring more robust cross-platform compatibility.
  • Test files are now included in Conda builds for better package quality assurance.
• YOLO11 Model References in Docs & Docker 📦
  • All documentation, code examples, and Dockerfiles now use YOLO11 models as the default, promoting the latest advancements.
• Cleaner CoreML Handling 🧹
  • Removed outdated CoreML and protobuf workarounds, reflecting improved compatibility and a simpler codebase.

🎯 Purpose & Impact

• Faster, Portable Text Features
  • The new TorchScript MobileCLIP model allows for rapid, device-agnostic text encoding, benefiting custom class naming and vision-language applications on a wider range of hardware.
• Broader Dataset Compatibility
  • Native grayscale support reduces manual preprocessing and errors, expanding Ultralytics' usability in medical imaging, document analysis, and more.
• Stronger Reliability Across Devices
  • Enhanced CI ensures YOLO models work smoothly on edge devices like Raspberry Pi and in Conda-managed environments, giving users and developers greater confidence.
• Simpler, Up-to-Date Experience
  • By standardizing on YOLO11 and cleaning up legacy code, users get clearer guidance and a more modern, maintainable toolkit.

---

This update is recommended for all users, especially those working with custom text prompts, grayscale datasets, or deploying on diverse hardware platforms. 🚀

What's Changed

• Remove special CoreML<>protobuf handling by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20362
• Add tests directory to package for Conda builds by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20365
• Implement 3.12-compatible Raspberry Pi CI by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20361
• Dockerfile export Sony IMX with YOLO11 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20367
• Add Grayscale image training support by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20310
• ultralytics 8.3.118 Use new TorchScript MobileCLIP model by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20354

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.117...v8.3.118

🌟 Summary

Ultralytics 8.3.117 introduces a major update to TFLite model exports, switching from flatbuffers to a simpler JSON metadata format for improved Python 3.12+ compatibility. This release also brings enhancements to Docker support, distributed training, and model export workflows, along with several bug fixes and documentation improvements. 🚀

---

📊 Key Changes

• TFLite Metadata Overhaul: Replaces the TFLite Support package and flatbuffers-based metadata with a new JSON-based approach, making TFLite exports compatible with Python 3.12 and above.
• Backward Compatibility: Automatically falls back to the legacy metadata format for Python versions below 3.12.
• Docker & CI Improvements: Updates Dockerfiles and GitHub Actions workflows for better Python 3.12 support and ARM64 benchmarking.
• MobileCLIP Dependency Update: Switches to the Ultralytics-maintained MobileCLIP fork for easier installation and improved reliability.
• Distributed Training Fixes: Enhances YOLO distributed training stability, especially on Windows and with different PyTorch backends.
• CoreML Export Fix: Resolves box size issues when exporting models to CoreML.
• Prediction Robustness: Restores automatic class inference in detection models when metadata is missing, improving prediction reliability.
• Documentation Upgrade: Adds a step-by-step YouTube video tutorial to the Docker Quickstart guide for easier onboarding.
• Codebase Cleanup: Refines import statements for better maintainability and fewer import errors.

---

🎯 Purpose & Impact

• Future-Proof TFLite Exports: Ensures users can export and use TFLite models with the latest Python versions, reducing compatibility headaches.
• Smoother Installations: The new MobileCLIP fork and Docker updates make setup and deployment more reliable across platforms.
• Easier Distributed Training: Users training YOLO models on clusters or Windows systems will experience fewer errors and smoother runs.
• Broader Hardware Support: ARM64 benchmarking and Docker improvements benefit users on Jetson and other ARM devices.
• Improved User Experience: The new Docker video guide and restored prediction logic make it easier for both beginners and advanced users to get started and achieve reliable results.
• ⚠️ Breaking Change: TFLite metadata format has changed. If you rely on the old flatbuffers-based metadata, you may need to update your code or workflows.

---

This release is a significant step towards modern Python compatibility, easier deployments, and a more robust user experience across the Ultralytics ecosystem! 🎉

What's Changed

• Add https://youtu.be/IYWQZvtOy_Q to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20347
• Fix private local import issue by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20343
• Fix DDP with Gloo backend on Windows by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20330
• Use Ultralytics MobileCLIP fork for YOLOE by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20337
• Fix predictions if metadata is missing by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20353
• Add Benchmarks CI to Dockerfile-arm64 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/20236
• Fix box sizes for CoreML exports by @sidekickr in https://github.com/ultralytics/ultralytics/pull/20357
• ultralytics 8.3.117 Replace TFLite Support with JSON metadata by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/16413

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.116...v8.3.117

🌟 Summary

This release brings enhanced loss function customization for advanced model training, improved export options, clearer documentation, and several usability and performance upgrades across the Ultralytics ecosystem. 🚀

📊 Key Changes

• Customizable Loss Functions: Added gamma and alpha parameters to FocalLoss and VarifocalLoss, allowing users to better handle class imbalance and focus on hard-to-classify examples.
• YOLOE Model Naming Consistency: Updated all YOLOE references and documentation to use the latest YOLO11-based naming (e.g., yoloe-11s-seg.pt), making it easier to select and use the correct models.
• TorchScript Export Enhancement: Added support for exporting models in half-precision (half argument) for TorchScript, improving performance on compatible hardware.
• Improved Non-Max Suppression (NMS): Fixed class filtering logic in NMS to ensure more accurate and reliable detection results when filtering by class.
• Export Flexibility: Introduced explicit control over bounding box output format (xyxy vs. xywh) during export, reducing confusion and making integration with deployment environments smoother.
• Ultralytics Solutions Label Customization: Added options to show/hide labels and confidence scores in Solutions modules, with a unified label formatting method for consistent output.
• Docker Image Update: Upgraded the base Docker image to PyTorch 2.7.0 for better compatibility and performance.
• Documentation Improvements: Refined docstrings, clarified return types, and updated code examples for better developer experience.
• Security Upgrade: Set GitHub workflow permissions to read-only for improved CI/CD security.

🎯 Purpose & Impact

• Greater Training Flexibility: Fine-tune loss function behavior to address challenging datasets, leading to potentially better model accuracy and robustness.
• Easier Model Selection: Consistent YOLOE naming reduces user confusion and errors in both code and documentation.
• Faster, Leaner Exports: Half-precision TorchScript exports enable faster inference and lower memory usage, especially on modern GPUs.
• More Reliable Detections: Improved NMS and class filtering logic ensure users get accurate results when working with specific object classes.
• Seamless Deployment: Explicit bounding box format control and Docker updates make it easier to deploy models in diverse environments.
• Customizable Visual Outputs: Solutions users can now tailor label and confidence display to their needs, enhancing presentation and clarity.
• Better Developer Experience: Improved documentation and code clarity help both new and experienced users work more efficiently.
• Enhanced Security: Workflow permission changes follow best practices, reducing risk in automated processes.

---

This update is packed with improvements for both model developers and end users, making Ultralytics tools more powerful, flexible, and user-friendly. 💡✨

What's Changed

• Update Tuple return type annotations in docstrings by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20274
• Fix YOLOE usage in model.py by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20321
• Fix non_max_suppression when applying classes argument by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20322
• Add half argument to Torchscript export allowlist by @seungjlee in https://github.com/ultralytics/ultralytics/pull/20313
• Add show_conf and show_labels for Solutions by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20282
• FROM pytorch/pytorch:2.7.0-cuda12.6-cudnn9-runtime by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20335
• Remove redundant coordinate conversions for nms=True by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20334
• Add workflow-level CI permissions for security by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20336
• ultralytics 8.3.116 RTDETR VarifocalLoss gamma and alpha parameterization by @pow3rpi in https://github.com/ultralytics/ultralytics/pull/20292

New Contributors

• @seungjlee made their first contribution in https://github.com/ultralytics/ultralytics/pull/20313
• @pow3rpi made their first contribution in https://github.com/ultralytics/ultralytics/pull/20292

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.115...v8.3.116

🌟 Summary

Ultralytics 8.3.115 brings improved TIFF image support, enhanced documentation, better data loading diagnostics, and fixes for distributed training—making the platform more robust and user-friendly. 🚀

📊 Key Changes

• 🖼️ TIFF Image Support: Improved handling of TIFF/TIF images, especially single-frame RGB files, ensuring reliable training and prediction with these formats.
• ⚡ Data Loading Diagnostics: Added warnings for slow disk read speeds, helping users quickly identify and resolve data bottlenecks.
• 🛠️ Distributed Training Fix: Corrected device assignment in multi-node, multi-GPU setups for smoother distributed training.
• 📚 Documentation Enhancements:
  • Updated YOLOE model usage examples to reflect the latest YOLO11 models and best practices.
  • Expanded Google-style docstring guidelines for clearer, more consistent code documentation.
  • Refined code formatting in documentation for improved readability.

🎯 Purpose & Impact

• Broader Image Compatibility: Users can now confidently use TIFF images (including RGB) in their projects without errors, supporting a wider range of datasets and workflows.
• Faster Troubleshooting: New disk speed checks make it easier to spot and fix slow data loading, leading to smoother training and evaluation.
• Seamless Distributed Training: Fixes prevent errors in advanced multi-node GPU setups, ensuring reliable scaling for large projects.
• Better Learning & Contribution: Improved documentation and examples help both new and experienced users get started quickly and contribute high-quality code.
• Backward Compatibility: All updates are designed to work smoothly with existing projects, minimizing disruption.

✨ This release is all about making Ultralytics more reliable, accessible, and efficient for everyone!

What's Changed

• Remove nested spans form Mkdocstrings HTML by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20293
• Add additional Google-style docstring examples by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20309
• Add slow disk read speed warning by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20303
• Fix usage in yoloe.md FAQ section by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20307
• Bug Fix: Incorrect Use of RANK in Distributed Context Manager by @SimoneCaldarella in https://github.com/ultralytics/ultralytics/pull/20312
• ultralytics 8.3.115 TIFF image RGB training and prediction support by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20301

New Contributors

• @SimoneCaldarella made their first contribution in https://github.com/ultralytics/ultralytics/pull/20312

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.114...v8.3.115

🌟 Summary

This release introduces advanced tracker re-identification (ReID) for object tracking, along with several improvements to documentation, dataset handling, and model export reliability. 🚀🔍

📊 Key Changes

• Tracker Re-Identification (ReID) Support:
  • Added native feature-based and model-based ReID to the BoTSORT tracker, enabling recovery of lost object tracks.
  • Default ReID mode is now auto, using YOLO model features when available.
  • Enhanced feature extraction and association logic for more robust multi-object tracking.
• Documentation Enhancements:
  • Cleaned up code block line numbers and improved guide titles for easier navigation and readability.
  • Added clear instructions for preparing multispectral TIFF images and updated dataset size info for DOTA8 Multispectral.
• Model Export & Metadata Fixes:
  • Ensured the "channels" field is always treated as an integer during MNN export, improving compatibility.
  • Fixed logging for INT8 export calibration to display the correct data path.

🎯 Purpose & Impact

• Smarter Object Tracking:
  • The new ReID feature allows the tracker to re-identify and reconnect lost objects, making tracking more reliable in challenging scenarios (e.g., occlusions or crowded scenes).
  • Users can benefit from improved tracking accuracy with minimal configuration—especially when using YOLO models.
• Better Documentation Experience:
  • Easier-to-read docs and clearer dataset instructions help both new and advanced users get started and troubleshoot more efficiently.
• Smoother Model Deployment:
  • Improved metadata handling and clearer export logs reduce errors and confusion when exporting and deploying models, especially in diverse environments.

Overall, this update brings smarter tracking, a smoother user experience, and more robust model handling to Ultralytics users! 🦾📈

What's Changed

• Remove href attributes from code line numbers in Docs code blocks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20286
• Parse channels as int in metadata by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20288
• Restore missing f-string by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20280
• Refactor code line number links to use <span> tags by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20290
• Update coco8-multispectral docs by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20272
• Update dota8-multispectral.yaml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20270
• ultralytics 8.3.114 New tracker re-identification (ReID) of lost tracks by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20192

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.113...v8.3.114

🌟 Summary

Ultralytics 8.3.113 introduces support for the DOTA8-Multispectral dataset, enhanced multi-channel image handling, improved Intel device selection for OpenVINO, and several usability and visualization upgrades across the platform. 🚁🌈💡

📊 Key Changes

• DOTA8-Multispectral Dataset Support: Added a new dataset configuration for 10-channel multispectral images, enabling advanced aerial and remote sensing applications.
• Improved Multispectral Conversion: Updated the convert_to_multispectral function and examples for seamless 10-channel dataset creation.
• Intel Device Selection for OpenVINO: Users can now specify Intel hardware (GPU, NPU, CPU) for inference, with automatic device checks and helpful warnings.
• Flexible Validation: Added a classes argument to validation, allowing users to evaluate models on specific object classes.
• Robust Multi-Channel Image Handling: Enhanced image processing logic to correctly handle grayscale, RGB, and multi-channel images without errors.
• SAHI OBB Video Inference Improvements: Easier device selection, new slice size options, and clearer documentation for video inference with SAHI and YOLO11.
• Region Counting Visualization: Improved color-based tracking and label readability for region counting tasks.
• Ultralytics Documentation Widget Upgrades: Modernized chat/search support widget with better branding, dark mode, and quick-access help links.

🎯 Purpose & Impact

• Advanced Research Ready: Enables training and evaluation on multispectral datasets, unlocking new use cases in aerial, satellite, and scientific imaging. 🛰️
• Greater Flexibility: Users can fine-tune validation and inference to their needs—whether focusing on specific classes or leveraging different Intel hardware for optimal performance.
• Improved Reliability: More robust handling of diverse image types prevents crashes and ensures accurate results across workflows.
• Enhanced Usability: Easier customization and clearer documentation make advanced features more accessible to all users.
• Better Support Experience: The upgraded documentation widget ensures users can quickly get help, ask questions, and access community resources.

This release empowers both researchers and practitioners with new dataset capabilities, smarter hardware utilization, and a smoother user experience throughout the Ultralytics ecosystem! 🚀

What's Changed

• Add classes to validation args by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20250
• Intel device support by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/20203
• Update extra.js with cxkit-js by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20266
• Update extra.js with dark mode fixes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20268
• Eliminate channel resorting for multi-channel images by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20263
• Add OBB support for SAHI by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20256
• Increase region count display size and add color-based visualization by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20235
• ultralytics 8.3.113 New DOTA8-Multispectral dataset by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20269

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.112...v8.3.113

🌟 Summary

This release introduces full support for multispectral (multi-channel) images in the Ultralytics ecosystem, allowing YOLO models to train, validate, predict, and export with images containing more than 3 channels (e.g., 10-channel multispectral data). 🌈

---

📊 Key Changes

• Multispectral Image Support:

  • Added robust handling for images with any number of channels (not just RGB) across all YOLO tasks (detection, segmentation, pose, classification, etc.).
  • Introduced a channels field in dataset configuration files to specify the number of image channels.
  • Updated data loaders, caching, and image reading utilities to correctly process multi-channel images, including TIFF support.
  • All model initialization, training, validation, and export routines now dynamically adapt to the dataset's channel count.

• New COCO8-Multispectral Dataset:

  • Added a 10-channel multispectral version of the COCO8 dataset for rapid testing and experimentation.
  • Provided a utility to convert standard RGB images to multispectral format via interpolation.
  • Comprehensive documentation and usage examples for multispectral datasets.

• Augmentation & Preprocessing Improvements:

  • Data augmentations and transformations now intelligently apply only to compatible channel configurations.
  • Enhanced plotting and visualization to handle multi-channel data gracefully.

• Other Notable Updates:

  • Improved MobileSAM documentation for clarity and easier comparison with YOLO models.
  • Updated callbacks documentation with a new, more relevant YouTube tutorial.
  • Enhanced dataset splitting utilities and documentation for classification tasks.
  • Standardized logging for warnings and errors, making messages clearer and more consistent.
  • Improved test coverage for ARM64 systems and multispectral workflows.

---

🎯 Purpose & Impact

• Unlocks Advanced Use Cases:

  • Enables researchers and practitioners to work with multispectral and hyperspectral imagery (e.g., satellite, medical, or scientific images) directly in Ultralytics and YOLO models.
  • Facilitates new applications in agriculture, remote sensing, and any domain requiring spectral analysis beyond standard RGB.

• Seamless Integration:

  • Multispectral support is built-in and automatic—no need for custom code or workarounds.
  • All core YOLO workflows (training, validation, prediction, export) now support multi-channel data out of the box.

• Enhanced Experimentation:

  • The new COCO8-Multispectral dataset provides a quick, lightweight way to test multispectral pipelines and debug models.
  • Utility functions make it easy to convert existing datasets to multispectral format for experimentation.

• Improved Usability & Documentation:

  • Clearer docs, better logging, and more robust dataset handling make the platform easier to use for both new and advanced users.
  • ARM64 and cross-platform improvements ensure broader compatibility.

---

In summary:
This update is a major step forward for users needing advanced image analysis, making Ultralytics and YOLO models more versatile and ready for real-world, multi-channel data challenges. 🚀

What's Changed

• Update MobileSAM documentation by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20219
• Add https://youtu.be/ENQXiK7HF5o to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20218
• New COCO8-Multispectral dataset by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20221
• Add Tests CI after building Arm64 Docker images by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19672
• RTDETR: Remove unused preprocess_batch function by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20215
• Fix stale workflow by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20227
• Update extra.js with 0.3.20/dist/embed.min.js by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20230
• Allow validation with rect for dynamic models by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20232
• Scope tensorboard.SummaryWriter import by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20220
• Fix test_solutions.py Streamlit test by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20233
• Classify results background transparency by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20239
• Fix Annotator tuples>lists by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20242
• Add PIL results.plot(save=True) test by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20241
• Refactor autosplit and implement split_classify_dataset by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20245
• Integrate "WARNING ⚠️" and "ERROR ❌" prefixes for LOGGER by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20246
• Update build_reference.py for automatic mkdocs.yml updates by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20247
• ultralytics 8.3.12 New YOLO Multispectral Image Support by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20223

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.111...v8.3.112

🌟 Summary

This release brings significant efficiency improvements to YOLOv10 models, enhances logging and integration options, and streamlines user experience in Ultralytics tutorials and workflows. 🚀✨

📊 Key Changes

• YOLOv10 Model Optimization: The fuse() method now removes the "one2many" detection head when not needed, reducing model size and computation for inference.
• Faster, Lighter YOLOv10: After fusion, YOLOv10n drops from 125 to 102 layers and from 2.76M to 2.30M parameters, making it faster and more efficient for deployment.
• TensorBoard Logging Default Changed: TensorBoard is now disabled by default for faster training startup; users can re-enable it as needed.
• Tutorial Notebook Improvements: The YOLO11 Colab notebook has clearer instructions, updated visuals, and improved links to documentation and community resources.
• Better Integration Options: Users can now easily select between TensorBoard and Weights & Biases for experiment tracking.
• Cleaner Output: Suppressed unnecessary warnings from the Albumentations library for a smoother notebook experience.
• TensorRT Export Fix: Improved support for exporting YOLO-World models to TensorRT by handling dynamic shapes more robustly.
• Stronger Link Checking: Documentation workflows now fail if broken links are found, ensuring higher quality docs.

🎯 Purpose & Impact

• Faster Inference & Smaller Models: The YOLOv10 optimization means users get quicker predictions and reduced resource usage, especially important for edge devices and production deployments.
• Streamlined Training: Disabling TensorBoard by default speeds up training for users who don’t need logging, while still allowing easy opt-in for those who do.
• Improved Usability: The revamped tutorial and cleaner output make it easier for both beginners and experienced users to get started and stay productive.
• Robust Integrations: Enhanced support for popular experiment tracking tools and export formats broadens compatibility and flexibility for diverse workflows.
• Higher Documentation Quality: Automated link checking helps ensure users always have access to up-to-date and working resources.

---

This update is all about making YOLO models faster, easier to use, and more reliable—whether you're deploying to production or just getting started! 🚦📈

What's Changed

• Explicit shape handling for dynamic YOLO-World exports to TensorRT by @laugh12321 in https://github.com/ultralytics/ultralytics/pull/20205
• Update links.yml to ignore TIMEOUT issues by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20210
• Default tensorboard=False for speed by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20213
• ultralytics 8.3.111 YOLOv10 skip one2many head when fused by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20193

New Contributors

• @laugh12321 made their first contribution in https://github.com/ultralytics/ultralytics/pull/20205

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.110...v8.3.111

🌟 Summary

Ultralytics 8.3.110 introduces automatic dataset file access speed checks, enhanced documentation (especially for YOLOE and object counting with YOLO11), improved reliability in data handling, and includes helpful scripts directly in the package. 🚀

📊 Key Changes

• Dataset Speed Checks: Added a new feature that measures and logs how quickly your dataset files can be accessed, warning you if your storage is slow.
• YOLOE Documentation Upgrade: Expanded and clarified YOLOE docs, including step-by-step guides for text and visual prompting, and introduced prompt-free model usage.
• Object Counting Guide Update: Refreshed the object counting tutorial with a new video focused on YOLO11 and real-time counting.
• Shell Scripts Included: All .sh scripts are now bundled with the Ultralytics package, making automation and setup easier.
• Robust File Hashing: Improved file hashing to handle missing or inaccessible files gracefully, reducing potential errors.
• Documentation & Code Cleanups: Fixed links, improved comments, and made minor refinements throughout the codebase.

🎯 Purpose & Impact

• Performance Awareness: Users are now alerted if their dataset storage is slow, helping prevent bottlenecks during training or inference. 🏎️
• Easier Learning: Updated guides and videos make it simpler for users to get started with object counting and advanced YOLOE features. 📚
• Smoother Automation: Including shell scripts directly in the package streamlines setup and custom workflows. ⚙️
• More Reliable Data Handling: Improved hashing and error handling mean fewer interruptions and a smoother experience, especially with large or remote datasets.
• Clearer Documentation: Enhanced docs empower both new and experienced users to make the most of Ultralytics tools. 📝

This release focuses on making Ultralytics faster, easier to use, and more robust for everyone!

What's Changed

• Add https://youtu.be/vKcD44GkSF8 to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20190
• Include bash scripts in installation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20186
• Improve hashing speed by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20184
• Improve clarity of YOLOE prediction docs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20201
• ultralytics 8.3.110 New dataset file access speed checks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20197

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.109...v8.3.110

🌟 Summary

This release brings improved control and reliability to model evaluation, training, and export, with a special focus on RT-DETR validation and memory management. 🚀

📊 Key Changes

• RT-DETR Validator Update: Added a confidence threshold filter to the RT-DETR model validator, so only predictions above a user-defined confidence level are considered during evaluation.
• CoreML Export Simplification: Streamlined CoreML export for classification models, ensuring smoother and more consistent exports.
• AutoBatch Logic Refinement: Improved automatic batch size calculation for training, making it more accurate and robust.
• Better GPU Memory Management: Lowered the threshold for clearing GPU memory during training from 90% to 50%, reducing out-of-memory risks.
• Testing Improvements: Solution tests are now conditionally skipped on Linux (Python 3.11) and Raspberry Pi to avoid known issues.
• Documentation & CI Enhancements: Fixed a broken link in the VS Code integration docs and improved CI reliability for link checking.

🎯 Purpose & Impact

• More Accurate Evaluation: The RT-DETR confidence filter ensures evaluation metrics better reflect real-world performance, giving users greater control and trust in their results.
• Smoother Model Export: Classification models exported to CoreML will now work more reliably, reducing errors and setup confusion.
• Stable Training Experience: Improved batch size selection and earlier GPU memory clearing help prevent crashes and slowdowns, especially on devices with limited resources.
• Reliable Testing & Documentation: Skipping problematic tests and fixing documentation links make the development experience smoother for both users and contributors.

Overall, this update makes Ultralytics models easier to use, more reliable, and better aligned with real-world needs. 🌟

What's Changed

• Remove unreachable CoreML Classify model logic by @rromanchuk in https://github.com/ultralytics/ultralytics/pull/20167
• AutoBatch: Revert polynomial fitting logic by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20173
• Reduce trigger threshold to clear GPU memory to 50% by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20179
• Update links.yml with set -o pipefail by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20181
• Enable Solutions tests, except on Linux (Python 3.11) and Raspberry PI due to --slow test errors by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20174
• ultralytics 8.3.109 Add conf filter to RT-DETR validator by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20175

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.108...v8.3.109

🌟 Summary

The release of v8.3.108 introduces enhanced support for YOLO11 models, particularly focusing on Sony IMX500 integration for edge AI applications, alongside various improvements in code consistency, documentation, and performance.

📊 Key Changes

• Sony IMX500 Export for YOLO11:
  • Updated export logic and documentation to support YOLO11 models.
  • Improved validation for supported models during export.
  • Benchmarked YOLO11n on IMX500, showing improved accuracy with a slight increase in inference time.
• Code Enhancements:
  • Deprecated the crop_fraction parameter for classification transforms, simplifying the codebase.
  • Replaced deprecated NumPy functions with version-aware alternatives for future compatibility.
  • Optimized pose angle calculations using Python's math module.
• Documentation Updates:
  • Added a comprehensive guide on YOLO data augmentation with visual examples and detailed explanations.
  • Embedded a YouTube tutorial in the project goals guide for better user engagement.
  • Improved formatting and readability across various documentation files.
• New Example:
  • Introduced an interactive object tracking UI using YOLO models, enabling real-time tracking with features like click-to-track and visual overlays.

🎯 Purpose & Impact

• Enhanced Compatibility: YOLO11 models can now be exported and deployed on Sony IMX500 hardware, expanding their usability in edge AI scenarios like Raspberry Pi AI Cameras.
• Improved User Experience: The new data augmentation guide and interactive tracking UI example provide practical tools and insights for both beginners and advanced users.
• Streamlined Development: Codebase simplifications and compatibility updates ensure smoother workflows and future-proofing for developers.
• Better Learning Resources: Enhanced documentation and video tutorials make it easier for users to understand and implement YOLO models effectively. 🚀

What's Changed

• Fix Solutions warnings by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20127
• Add squeeze() in category_freq before indexing by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20131
• Add new example Ultralytics YOLO-Interactive-Tracking-UI by @alireza787b in https://github.com/ultralytics/ultralytics/pull/19854
• Disable Solutions test by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20137
• Improve headers, redirects and example READMEs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20145
• Add https://youtu.be/q1tXfShvbAw to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20149
• Add new YOLO Data Augmentation guide by @picsalex in https://github.com/ultralytics/ultralytics/pull/20124
• Deprecate crop_fraction parameter by @fcakyon in https://github.com/ultralytics/ultralytics/pull/20146
• Fix Comet error during COCO validation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20153
• Fix hub/pro.md docs page indentation issues by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20157
• Align code examples in yolo-data-augmentation page by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20158
• ultralytics 8.3.108 YOLO11 Sony IMX export by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/19946

New Contributors

• @alireza787b made their first contribution in https://github.com/ultralytics/ultralytics/pull/19854

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.107...v8.3.108

🌟 Summary

The Ultralytics 8.3.107 release introduces key improvements for model compatibility, export functionality, and testing efficiency. 🚀

---

📊 Key Changes

• Rockchip RKNN Fix: Resolved file handling issues in torch_to_mnn by ensuring paths are properly converted to strings.
• OpenVINO Compatibility: Updated OpenVINO version requirements to avoid macOS-specific issues and reverted CI runners to macOS 14 for stability.
• Export Enhancements: Added standalone functions for exporting PyTorch models to ONNX and TensorRT formats, simplifying workflows for benchmarking and deployment.
• Ray Tune Improvements: Introduced shorter, consistent trial names and directories for better organization during hyperparameter tuning.
• BatchNorm Update: Ensured data type consistency in convolution and batch normalization fusion for improved precision.
• Testing Optimization: Streamlined video download logic in solution tests, reducing unnecessary operations.

---

🎯 Purpose & Impact

• 🛠️ Improved Compatibility: Enhanced support for RKNN models and resolved OpenVINO/macOS issues, ensuring smoother workflows across platforms.
• 📂 Simplified Export: New export utilities make it easier to convert and benchmark models in ONNX and TensorRT formats.
• 🔍 Better Debugging: Clearer trial naming in Ray Tune simplifies tracking and debugging during tuning processes.
• ⚡ Enhanced Precision: Updates to batch normalization fusion improve model stability and performance, especially for mixed-precision setups.
• 🚀 Efficient Testing: Optimized test logic saves time and resources, improving the development experience.

This release focuses on reliability, usability, and compatibility, making it a valuable update for developers and researchers alike. 🌟

What's Changed

• Pin openvino<2025.0.0 to fix OpenVINO export issue on macOS by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20112
• Use shorter trial names for Ray Tune by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20109
• Revert CI runners to macOS 14 to avoid OpenVINO bug on macOS 15.4 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20126
• Enable solutions tests for each PR by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20110
• Add missing batchnorm dtype arg by @kstreee-furiosa in https://github.com/ultralytics/ultralytics/pull/20096
• Add standalone export functions for onnx and tensorrt formats by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20074
• ultralytics 8.3.107 Rockchip RKNN Autobackend fix by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20125

New Contributors

• @kstreee-furiosa made their first contribution in https://github.com/ultralytics/ultralytics/pull/20096

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.106...v8.3.107

🌟 Summary

This release introduces enhanced CoreML export capabilities for iOS14+ devices, along with several optimizations and fixes to improve user experience and compatibility. 🚀

---

📊 Key Changes

• iOS14+ CoreML Export Support:

  • Added support for exporting models in both mlmodel and mlpackage formats for iOS14+ using coremltools 8.x.
  • Removed the minimum iOS version parameter, allowing CoreML tools to automatically determine compatibility based on the model.
  • Adjusted NMS (Non-Maximum Suppression) export to align its specification version with the core model.

• Package Cleanup:

  • Excluded the top-level tests directory from the package distribution to avoid conflicts with user projects.

• Documentation Fix:

  • Updated a broken link in the Albumentations integration guide for better navigation.

• Tutorial Notebook Update:

  • Replaced !uv pip install ultralytics with %pip install ultralytics for improved compatibility in Jupyter environments.

---

🎯 Purpose & Impact

• CoreML Export Enhancements:

  • Enables seamless deployment of YOLO models on iOS14+ devices, broadening accessibility for Apple users. 🍎
  • Simplifies the export process by leveraging CoreML tools' automatic compatibility settings.

• Cleaner Package Distribution:

  • Reduces potential naming conflicts and ensures a more streamlined installation experience for developers.

• Improved Documentation:

  • Enhances user experience by providing accurate and reliable resources for Albumentations integration. 📚

• Better Tutorial Compatibility:

  • Ensures smoother execution of tutorials in diverse Jupyter environments, making it easier for users to get started. 😊

This update focuses on improving usability, compatibility, and deployment flexibility, making it a valuable upgrade for developers and users alike. 🌟

What's Changed

• Remove top-level tests directory in package distribution by @esmalTT in https://github.com/ultralytics/ultralytics/pull/20079
• Fix broken link in albumentations.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20087
• Remove uv from tutorial.ipynb by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20106
• ultralytics 8.3.106 iOS14+ CoreML mlmodel and mlpackage joint support by @sidekickr in https://github.com/ultralytics/ultralytics/pull/20100

New Contributors

• @esmalTT made their first contribution in https://github.com/ultralytics/ultralytics/pull/20079
• @sidekickr made their first contribution in https://github.com/ultralytics/ultralytics/pull/20100

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.105...v8.3.106

🌟 Summary

The YOLO 8.3.105 release focuses on simplifying workflows, enhancing flexibility, and improving usability across various features and integrations. Key updates include the removal of unused arguments, better export device handling, and improved object counting visuals.

📊 Key Changes

• Removed save_hybrid Argument: Eliminated this rarely used feature from validation workflows, configurations, and documentation.
• Export Device Flexibility: Added a device parameter to specify hardware (e.g., GPU, CPU, MPS) for exporting models across formats like ONNX, TensorRT, CoreML, and more.
• Enhanced Object Counting: Improved analytics display with a new margin parameter for better text background scaling and visual clarity.
• YOLOv8 MNN C++ Example: Added a new example showcasing YOLOv8 integration with the MNN framework in C++.
• YOLOE-PF Export Fix: Resolved dimension issues for YOLOE-PF model exports, ensuring compatibility and accuracy.

🎯 Purpose & Impact

• Simplified Validation: Removing save_hybrid reduces confusion and prevents incorrect mAP calculations, streamlining the validation process.
• Improved Deployment: The device parameter allows users to optimize model exports for specific hardware, enhancing compatibility and performance, especially for edge devices.
• Better Analytics Display: The refined object counting visuals make it easier to interpret results in real-time applications.
• Broader Framework Support: The MNN C++ example expands YOLOv8's usability for developers working with lightweight frameworks.
• Export Reliability: Fixes to YOLOE-PF exports ensure smoother workflows and accurate model outputs.

This release enhances both the developer experience and the robustness of YOLO integrations, making it more accessible and efficient for diverse use cases. 🚀

What's Changed

• Add support to adjust text background size in ObjectCounter with line_width by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20073
• Update MNN example README by @vnk8071 in https://github.com/ultralytics/ultralytics/pull/20065
• Add device argument to all exports by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/20009
• Fix YOLOE-PF export dimension by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20072
• ultralytics 8.3.105 Remove unused save_hybrid argument by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20067

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.104...v8.3.105

🌟 Summary

This release (8.3.104) focuses on improving error handling, model validation, export flexibility, and documentation clarity for a smoother user experience. 🚀

---

📊 Key Changes

• Error Fix in YOLOE Predictions: Resolved an issue where NumPy arrays caused ambiguous truth value errors during predictions.
• Improved YOLOE Validation: Added support for loading models directly from file paths during validation.
• IMX500 Export Enhancement: Introduced a device parameter to allow GPU or CPU selection for faster and more efficient exports.
• Documentation Updates: Fixed outdated links and improved clarity in YOLOv8 examples for MNN and ONNXRuntime integrations.

---

🎯 Purpose & Impact

• Error-Free Predictions: Ensures compatibility with all source types, including NumPy arrays, making predictions more robust and reliable. 🛡️
• Streamlined Validation: Simplifies workflows by enabling direct model loading from file paths, saving time and effort. ⏱️
• Export Flexibility: Allows users to optimize IMX500 exports by selecting the appropriate hardware (GPU or CPU), improving performance and adaptability. ⚡
• Enhanced Documentation: Reduces confusion and improves navigation for developers, fostering a better user experience and encouraging contributions. 📚

This update ensures smoother operations across various use cases, from model predictions to exports and documentation, benefiting both new and experienced users. 🌟

What's Changed

• YOLOE: Fix validation when model loading from local file by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20050
• Add device argument to Sony imx export by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19768
• Fix YOLOv8-ONNXRuntime example links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20059
• ultralytics 8.3.104 YOLOE explicit source is not None check by @JShengP in https://github.com/ultralytics/ultralytics/pull/20046

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.103...v8.3.104

🌟 Summary

The v8.3.103 release introduces a significant enhancement to hyperparameter tuning with the new RayTune resume=True functionality, alongside other updates improving model usability, stability, and documentation.

📊 Key Changes

• RayTune Resumable Tuning:
  • Added resume=True functionality for RayTune, allowing interrupted hyperparameter tuning sessions to be resumed seamlessly.
  • Automatically detects previous tuning runs and resumes from the last iteration.
  • Improved logging and directory handling for better user experience during tuning.
• Instance Segmentation Enhancements:
  • Refactored the InstanceSegmentation class to streamline mask plotting and annotation.
  • Introduced new configuration options (show_conf, show_labels, show_boxes) for better control over displayed annotations.
• CoreML Updates:
  • Increased CoreML specificationVersion from 5 to 9 for enhanced compatibility.
  • Lowered the minimum deployment target to iOS 15, broadening device support.
• Dependency Fixes:
  • Pinned shapely to <2.1.0 and streamlit to <1.44.0 to ensure compatibility and stability.
• Documentation Improvements:
  • Updated various guides and examples for better clarity and usability.
  • Enhanced FLOPs calculation logic and added detailed annotations for improved developer experience.

🎯 Purpose & Impact

• Efficient Hyperparameter Tuning:
  • The new RayTune resume=True feature saves time and resources by allowing users to continue interrupted tuning sessions without starting over.
  • Simplifies workflows for users conducting extensive hyperparameter searches.
• Improved Visualization:
  • Enhanced instance segmentation plotting provides clearer and more customizable annotations, making results easier to interpret.
• Broader Compatibility:
  • CoreML updates ensure better support for Apple devices, including older iOS versions, and improve export functionality.
• Stability and Reliability:
  • Dependency fixes prevent potential issues with future library updates, ensuring smoother user experiences.
• Developer-Friendly Enhancements:
  • Improved documentation and code annotations make it easier for developers to understand and utilize the tools effectively.

This release focuses on making hyperparameter tuning more robust, improving visualization capabilities, and ensuring compatibility across platforms, all while enhancing the overall user and developer experience. 🚀

What's Changed

• YOLOE: Add assert in set_classes to check prompt-free models by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20024
• Improve instance-segmentation mask plotting by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20025
• Truncate long Docs meta titles by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20030
• Fix solutions tests by pinning shapely<2.1.0 and streamlit<1.44.0 by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20028
• Add example MNN YOLOv8 using C++ by @vnk8071 in https://github.com/ultralytics/ultralytics/pull/19948
• Fix inconsistent value of max F1 conf from F1_curve.png caused by f value for smooth function in plot_mc_curve by @fazrigading in https://github.com/ultralytics/ultralytics/pull/20035
• Improve FLOPs docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20040
• Increase CoreML specificationVersion 5 -> 9 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20041
• CoreML minimum_deployment_target=ct.target.iOS15 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20042
• ultralytics 8.3.103 New RayTune resume=True functionality by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/20037

New Contributors

• @vnk8071 made their first contribution in https://github.com/ultralytics/ultralytics/pull/19948
• @fazrigading made their first contribution in https://github.com/ultralytics/ultralytics/pull/20035

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.102...v8.3.103

🌟 Summary

The YOLOE module has undergone a significant refactor in version 8.3.102, introducing new neural network modules and reorganizing the codebase for improved modularity and functionality. This update enhances the framework's flexibility and performance for advanced neural network operations.

📊 Key Changes

• New Modules Introduced:
  • SwiGLUFFN: A feed-forward network tailored for transformer-based architectures.
  • Residual: A module enabling residual connections to enhance learning efficiency.
  • SAVPE: A Spatial-Aware Visual Prompt Embedding module for feature enhancement.
• Code Reorganization:
  • Moved SwiGLUFFN, Residual, and SAVPE from head.py to block.py for better modularity.
  • Updated related documentation to reflect these structural changes.
• Version Update: Incremented version from 8.3.101 to 8.3.102.

🎯 Purpose & Impact

• Purpose:
  • Enhance the framework's capabilities with new modules designed for advanced neural network tasks.
  • Improve code maintainability and modularity by grouping related components logically.
• Impact:
  • Developers gain access to more powerful tools for building and customizing neural networks.
  • The reorganization simplifies the codebase, making it easier to extend and maintain.
  • These updates pave the way for more efficient and scalable model architectures, benefiting both research and production use cases. 🚀

What's Changed

• Remove verbose build_text_model log by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19998
• Update GNU links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20001
• Add fraction argument support for export dataloader by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/19952
• Refactor tablesort.js by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20002
• Remove MkDocs redirects by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20003
• Remove all gh-pages files except vercel.json by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20006
• Fix docs.yml vercel.json handling by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20007
• Parameterize solutions tests by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19961
• YOLOE: Remove unused argument max_det by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20011
• YOLOE: Bypass source check when source is None by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20012
• Update models/ docstrings by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/20017
• Add https://youtu.be/HMOoM2NwFIQ to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/20020
• Fix scaling ratio calculation in ONNX example by @MrBlackBigWhiteSheep in https://github.com/ultralytics/ultralytics/pull/20016
• ultralytics 8.3.102 YOLOE module refactor by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/20010

New Contributors

• @MrBlackBigWhiteSheep made their first contribution in https://github.com/ultralytics/ultralytics/pull/20016

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.101...v8.3.102

🌟 Summary

The YOLOE model has been enhanced with improved video/stream handling, better rectangular inference options, and updated documentation for a smoother user experience. 🚀

---

📊 Key Changes

• Visual Prompt for Videos: Automatically selects the first frame of a video/stream as the reference image for visual prompts, or allows users to specify a frame. 🎥
• Rectangular Inference Option: Introduced a rect parameter to control rectangular padding during predictions, offering more flexibility.
• Documentation Updates: Fixed errors in YOLOE examples, clarified bounding box formatting, and removed outdated integration notes. 📚
• Model Export Improvements: Enhanced handling of class positional embeddings for better compatibility during model export.
• Contributor Recognition: Added new contributors to the project metadata, fostering community engagement. 🤝

---

🎯 Purpose & Impact

• Simplified Video Handling: Automates reference frame selection for video/stream inputs, saving users time and effort when applying visual prompts.
• Enhanced Prediction Flexibility: The rect parameter allows users to toggle rectangular padding, improving adaptability for different use cases.
• Improved Documentation: Clearer examples and updated notes reduce confusion, making it easier for users to implement YOLOE in their projects.
• Better Export Compatibility: Ensures smoother model operations during export, particularly for advanced use cases.
• Community Engagement: Recognizing contributors strengthens the Ultralytics community and highlights collaboration.

These updates collectively enhance usability, flexibility, and performance, ensuring a more seamless experience for developers and users alike. 🚀

What's Changed

• YOLOE: Fix visual prompt usage in docs by @JShengP in https://github.com/ultralytics/ultralytics/pull/19981
• YOLOE: Fix simplify with dynamic export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19988
• Update yoloe.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19982
• Enable rect option to predict mode by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19994
• YOLOE: Fix Linear Probing example by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19985
• ultralytics 8.3.101 YOLOE visual prompt inference fix for video sources by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19959

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.100...v8.3.101

🌟 Summary

The v8.3.100 release introduces enhanced support for PaddlePaddle, ensuring compatibility with the latest versions and improving model export functionality. Additionally, it includes updates to documentation, pre-trained model links, and usability improvements for YOLO11 and YOLOE models.

---

📊 Key Changes

• PaddlePaddle Integration:
  • Updated support for PaddlePaddle >=3.0.0 in both GPU and CPU environments.
  • Improved model export handling for .json and .pdiparams files.
  • Disabled OBB model inference for Paddle due to a silent error causing mAP drops (issue raised with PaddlePaddle).
• Documentation Enhancements:
  • Added a new video tutorial for YOLO11 with NVIDIA DeepStream.
  • Updated guides for YOLO11 deployment on Jetson devices.
  • Introduced detailed references for YOLOE's visual prompt capabilities and text encoding models like CLIP and MobileCLIP.
• Pre-trained Model Updates:
  • All pre-trained model links updated to v8.3.0 for YOLOv8, YOLOv9, YOLOv10, YOLO-NAS, YOLOE, and others.
• YOLOE Improvements:
  • Prevented unintended modifications to visual prompts during predictions.
  • Enhanced task management to automatically detect segmentation or detection tasks based on the predictor.

---

🎯 Purpose & Impact

• Purpose:
  • Ensure compatibility with the latest PaddlePaddle versions and improve export reliability.
  • Provide clearer, updated documentation for easier model deployment and usage.
  • Enhance YOLOE's stability and usability for workflows involving visual prompts and mixed tasks.
• Impact:
  • 🚀 Improved Compatibility: Users can now leverage PaddlePaddle >=3.0.0 for seamless integration.
  • 📚 Better Documentation: Simplifies deployment and training processes with updated guides and resources.
  • 🔧 Enhanced Usability: YOLOE's improvements reduce errors and streamline workflows for developers.
  • 🎥 User-Friendly Resources: New video tutorials make it easier for users to adopt YOLO11 on NVIDIA platforms.

This release ensures smoother workflows, better compatibility, and an overall improved user experience. 🌟

What's Changed

• Add https://youtu.be/hvGqrVT2wPg to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19943
• Update assets links by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19949
• Update DeepStream doc with YOLO11 support by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19958
• YOLOE: Prevent original prompts from being modified by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19963
• YOLOE: Preserve task type with refer_image by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19969
• ultralytics 8.3.100 New paddlepaddle>=3.0.0 with *.pdiparams by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19902

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.99...v8.3.100

🌟 Summary

The v8.3.99 release introduces YOLOE models, a groundbreaking addition to the YOLO family, enabling advanced open-vocabulary detection, segmentation, and visual/text prompt-based tasks. This update also includes enhancements to Docker compatibility, object tracking examples, repository mirroring workflows, and documentation improvements.

📊 Key Changes

• YOLOE Model Integration: Added YOLOE models for open-vocabulary detection and segmentation, supporting text and visual prompts.
• Prompt-Free Mode: Enabled training and inference without predefined prompts for broader application scenarios.
• Visual Prompt Embedding (SAVPE): Introduced spatial-aware visual prompt embedding for enhanced feature extraction.
• Docker Updates: Added Java Runtime Environment (JRE) and specific numpy version for Sony IMX model export compatibility.
• Object Tracking Enhancements: Improved YOLO11 tracking examples to handle edge cases and enhance visualization.
• Repository Mirroring: Streamlined GitHub-to-DagsHub mirroring workflow with manual triggering and improved flexibility.
• Documentation Overhaul: Updated image links, dataset guidance, and training instructions for better clarity and usability.

🎯 Purpose & Impact

• Enhanced Model Capabilities: YOLOE models expand YOLO's functionality to detect and segment objects beyond predefined categories, making it ideal for dynamic, real-world applications.
• Broader Use Cases: Prompt-free mode and SAVPE enable YOLOE to adapt to diverse scenarios, from autonomous systems to creative AI tasks.
• Improved Developer Experience: Docker updates simplify model export workflows, while tracking and documentation enhancements make implementation more robust and user-friendly.
• Streamlined Repository Management: Automated mirroring ensures consistent updates across platforms, saving time and improving accessibility for developers.

This release significantly elevates YOLO's versatility and usability, catering to both cutting-edge research and practical deployment scenarios. 🚀

What's Changed

• Update Dockerfile-python with JRE by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19925
• Update object tracking examples in docs by @ankanpy in https://github.com/ultralytics/ultralytics/pull/19861
• New Mirror Repository GitHub Action by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19846
• Fix Mirror Action by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19926
• Documentation refactor and improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19927
• ultralytics 8.3.99 New YOLOE Open-Vocabulary Models by @leonnil in https://github.com/ultralytics/ultralytics/pull/19775

New Contributors

• @leonnil made their first contribution in https://github.com/ultralytics/ultralytics/pull/19775

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.98...v8.3.99

🌟 Summary

The v8.3.98 release focuses on improving Java dependency handling for Sony IMX export, enhancing compatibility, and refining several other features across the Ultralytics ecosystem.

📊 Key Changes

• Java Dependency Update:
  • Updated Dockerfile to use default-jre instead of specific Java versions for Sony IMX export.
  • Enforced Java version 17 or higher for compatibility during the export process.
• Bug Fixes and Improvements:
  • Fixed color consistency in YOLO8 LibTorch C++ inference by ensuring padding uses RGB values.
  • Resolved a YOLO-NAS post-processing issue, aligning it with the broader Ultralytics framework.
  • Improved handling of keypoints in pose models to mark out-of-bounds keypoints as invisible instead of clipping.
• Documentation Enhancements:
  • Updated contribution guidelines for clarity and compliance with licensing.
  • Enhanced README files across examples and integrations for better usability.

🎯 Purpose & Impact

• 🛠 Simplified Java Setup: Using default-jre reduces installation complexity and ensures smoother Sony IMX export processes.
• ✅ Improved Accuracy: Fixes in padding and keypoint handling enhance model reliability and visual consistency.
• 📄 Better Documentation: Clearer guidelines and updated examples make it easier for developers to contribute and use Ultralytics tools effectively.
• 🚀 Framework Alignment: YOLO-NAS updates ensure better integration and maintainability within the Ultralytics ecosystem.

This release reflects Ultralytics' commitment to improving user experience, compatibility, and technical robustness across its tools and models.

What's Changed

• Fix C++ Example letterbox by @Q-qqq in https://github.com/ultralytics/ultralytics/pull/19900
• Fix NAS post-processing by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19882
• README enhancements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19906
• Fix keypoints being clipped rather than hidden by @HeadTriXz in https://github.com/ultralytics/ultralytics/pull/19921
• ultralytics 8.3.98 Sony IMX Java Runtime Environment>=17 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19905

New Contributors

• @Q-qqq made their first contribution in https://github.com/ultralytics/ultralytics/pull/19900
• @HeadTriXz made their first contribution in https://github.com/ultralytics/ultralytics/pull/19921

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.97...v8.3.98

🌟 Summary

The Ultralytics 8.3.97 release focuses on enhancing Sony IMX export capabilities, improving Docker usability, and ensuring compatibility with key dependencies. Additionally, it includes minor documentation and code quality updates for a better user and developer experience. 🚀

---

📊 Key Changes

• Sony IMX Export Enhancements:

  • Added Java (OpenJDK 17) to Docker for Sony IMX compatibility.
  • Updated dependencies for Sony IMX export, including model-compression-toolkit (>=2.3.0) and sony-custom-layers (>=0.3.0).
  • Adjusted Python requirements, specifying numpy==1.26.4 for Sony IMX workflows.
  • Introduced a new export command for Sony IMX models in Docker.

• Dependency Management:

  • Pinned paddlepaddle to versions below 3.0.0 to avoid compatibility issues.

• Documentation and Code Improvements:

  • Standardized license headers across files for clarity.
  • Fixed typos and improved consistency in example notebooks.
  • Enhanced type annotations and documentation for better code readability.

• Minor Adjustments:

  • Removed outdated redirects in documentation.
  • Improved HTML title handling for documentation pages.

---

🎯 Purpose & Impact

• Enhanced Sony IMX Support:

  • Streamlines Sony IMX export workflows by integrating necessary tools and dependencies directly into Docker.
  • Saves users time and effort in setting up their environment.

• Improved Stability:

  • Pinned paddlepaddle ensures smoother exports and inference workflows, avoiding potential bugs from breaking changes in newer versions.

• Better User Experience:

  • Simplified and standardized example notebooks make it easier for users to follow and implement workflows.
  • Documentation updates improve navigation and usability, helping users find information faster.

• Developer-Friendly Enhancements:

  • Consistent license headers and improved type annotations make the codebase easier to maintain and contribute to.

This release ensures a more robust and user-friendly experience for both end-users and developers, with a particular focus on expanding export capabilities and maintaining compatibility. 🌟

What's Changed

• Remove /SECURITY redirect in mkdocs.yml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19865
• Fix Example: -> Examples: by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19866
• Pin paddlepaddle<=3.0.0 to avoid bug in latest release by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19876
• Adopt H1 headers for Docs page HTML <title> by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19875
• Fix typo in notebooks by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19883
• ultralytics 8.3.97 Update Dockerfile-cpu Sony IMX export tools by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19765

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.96...v8.3.97

🌟 Summary

The Ultralytics 8.3.96 release focuses on simplifying Docker configurations, enhancing compatibility, and improving user experience for developers working with YOLO11 and related tools. 🚀

📊 Key Changes

• 🛠 Unified Docker Setup: Removed the Dockerfile-nvidia-cuda and consolidated configurations into the main Dockerfile.
• 📦 Preinstalled Key Libraries: Added tensorrt and onnxruntime-gpu to the Dockerfile for streamlined GPU workflows.
• 🔇 Cleaner Logs: Suppressed TensorFlow warnings for a more user-friendly experience.
• 🔄 Simplified PaddlePaddle Installation: Removed Python version constraints for easier compatibility.
• 📖 Documentation Updates:
  • Added overclocking instructions for Raspberry Pi 5 to boost YOLO11 performance.
  • Expanded Docker documentation with new Dockerfiles for NVIDIA Jetson devices and JupyterLab.
• ⚠️ Improved Version Warnings: Enhanced clarity in version mismatch messages to assist users in resolving issues quickly.

🎯 Purpose & Impact

• Streamlined Docker Management: By unifying and optimizing Docker configurations, users benefit from reduced complexity and easier maintenance.
• Enhanced GPU Workflows: Preinstalled libraries like tensorrt and onnxruntime-gpu save setup time and improve performance for GPU-based tasks.
• Improved Usability: Suppressing TensorFlow warnings and simplifying PaddlePaddle installation make the tools more accessible and user-friendly.
• Broader Hardware Support: New Dockerfiles and Raspberry Pi overclocking guidance enable better performance on diverse devices, from Jetson boards to Raspberry Pi.
• Clearer Error Handling: Improved version warnings reduce confusion, helping users troubleshoot faster.

This release ensures a smoother, more efficient experience for developers and users across various platforms and workflows. 🌟

What's Changed

• Fix latest-nvidia-cuda Docker tag by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19835
• Docker python3 > python symbolic links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19836
• Fix check_version() report with multiple constraints by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19840
• Update Raspberry Pi doc with overclock information by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19841
• Update Docker doc with missing Dockerfiles by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19842
• ultralytics 8.3.96 Preinstall tensorrt and onnxruntime-gpu in Dockerfile by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19845

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.95...v8.3.96

🌟 Summary

The Ultralytics 8.3.95 release introduces a CUDA-optimized Dockerfile for enhanced GPU support, updates dependencies, and improves documentation and usability across various tools and integrations.

📊 Key Changes

• New NVIDIA CUDA Dockerfile: Added Dockerfile-nvidia-cuda for optimized YOLO11 training and inference on GPUs.
• Dependency Updates: Upgraded PyTorch base image to 2.6.0-cuda12.6-cudnn9-runtime and CoreML compatibility to version 8.0.
• Improved Documentation: Enhanced formatting, added examples, and clarified CLI usage across guides.
• CLI Syntax Standardization: Ensured array inputs (e.g., classes, kpts) are enclosed in quotes for consistency.
• YOLOE Documentation Update: Added detailed tips for selecting detection modes.
• GitHub Banner Update: Redirects to the latest YOLO11 blog for better user engagement.

🎯 Purpose & Impact

• Enhanced GPU Performance: The new CUDA Dockerfile ensures seamless multi-GPU training and inference, benefiting users with high-performance hardware. 🚀
• Improved Compatibility: Updates to CoreML and PyTorch ensure support for the latest Python versions and GPU optimizations. 🖥️
• Better Usability: Documentation improvements and CLI standardization reduce user errors and enhance the overall experience. 📚
• Future-Proofing: Keeps the project aligned with the latest software advancements, ensuring long-term reliability and efficiency. 🌟

This release is a significant step forward for developers and users leveraging YOLO11 for advanced AI tasks.

What's Changed

• Add bash codeblock formatting by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19804
• Add quotes for solutions CLI arguments by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19806
• Fix tip display in YOLOE.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19813
• FROM pytorch/pytorch:2.6.0-cuda12.6-cudnn9-runtime by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19817
• Require coremltools>=8.0 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19819
• Add https://youtu.be/EXIpyYVEjoI to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19820
• Update GitHub banner redirect by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19828
• ultralytics 8.3.95 New dockerfile-nvidia-cuda FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu22.04 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19833

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.94...v8.3.95

🌟 Summary

The YOLO 8.3.94 release focuses on improving segmentation validation efficiency and enhancing documentation and code clarity across the project. 🚀

📊 Key Changes

• Segmentation Validation Improvements:
  • Increased the limit of plotted masks during validation to 50 for better visualization.
  • Added a warning when more than 50 masks are plotted to maintain speed and user awareness.
  • Updated maximum detection limit for plotting predictions to 50 for consistency. ⚡
• Documentation Enhancements:
  • Updated YOLOE documentation with corrected benchmarks and clarified threading and Ray Tune integration.
  • Fixed inaccuracies in the Tiger Pose dataset description.
• Code Standardization:
  • Replaced uppercase Dict, List, and Tuple with lowercase counterparts (dict, list, tuple) in type hints for modern Python compatibility.
• JavaScript Optimization:
  • Improved JS minification logic for cleaner and more efficient documentation handling.

🎯 Purpose & Impact

• Enhanced User Experience:
  • Improved segmentation validation plots provide more comprehensive insights into model performance while maintaining speed.
  • Clearer documentation ensures better understanding of YOLOE's capabilities and usage.
• Consistency and Modernization:
  • Standardized type hints align with Python best practices, improving code readability and maintainability.
• Efficiency Gains:
  • Optimized JavaScript handling reduces file sizes, enhancing documentation performance.

This update ensures a smoother and more efficient experience for developers and users working with YOLO models. 🌟

What's Changed

• Update YOLOE Docs page by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19776
• Use lowercase dict/list/tuple for type hint in docstrings by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19771
• Fix JS Minification by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19786
• Fix tiger-pose dataset description in docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19789
• ultralytics 8.3.94 Limit Segment val plots to 50 items/image for speed by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19792

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.93...v8.3.94

🌟 Summary

Release 8.3.93 introduces a critical fix for TorchScript models with Non-Maximum Suppression (NMS), along with several documentation and usability enhancements.

📊 Key Changes

• TorchScript Fix: Resolved a loading error for TorchScript models with NMS by ensuring torchvision is imported before model loading.
• YOLOE Documentation: Added comprehensive documentation for YOLOE, a new real-time open-vocabulary detection and segmentation model.
• Documentation Improvements:
  • Updated banner image to AVIF format for faster loading.
  • Minified HTML, CSS, and JS files for improved documentation performance.
  • Enhanced parameter descriptions for export, prediction, validation, and visualization tasks.
• New Features:
  • YOLOE supports text, visual, and internal prompts for detecting unseen object classes.
  • Added stream for memory-efficient video/image processing and txt_color for customizable annotation text colors.

🎯 Purpose & Impact

• TorchScript Fix: Ensures seamless loading and inference for TorchScript models, improving reliability for users deploying models in production. 🛠️
• YOLOE Introduction: Expands YOLO's capabilities to open-world scenarios, enabling detection of arbitrary objects with state-of-the-art performance while maintaining YOLO's speed. 🚀
• Documentation Enhancements:
  • Faster and more reliable documentation access due to optimized assets. ⚡
  • Clearer parameter descriptions improve usability for developers and users. 📚
• New Features:
  • YOLOE's versatility supports dynamic detection tasks, making it ideal for robotics, AR, and real-time applications. 🤖
  • Memory-efficient processing and customizable visuals enhance user experience and flexibility. 🎨

This release focuses on improving model robustness, introducing cutting-edge capabilities, and enhancing user experience through better documentation and new features.

What's Changed

• Update style.css announcement bar AVIF by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19760
• Fix Docs CSS, Javascript and HTML minification by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19763
• Update macros by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19764
• Create YOLOE Docs page by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19766
• ultralytics 8.3.93 Fix TorchScript Model with NMS loading error by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19747

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.92...v8.3.93

🌟 Summary

The YOLO11 v8.3.92 release focuses on enhancing single-class training functionality, improving compatibility for TensorFlow exports, and refining documentation. These updates aim to provide a smoother user experience and expanded capabilities for diverse use cases. 🚀

📊 Key Changes

• Single-Class Training Fix: Resolved a cache error during single-class training by updating label processing logic.
• TensorFlow Export Update: Added ai-edge-litert>=1.2.0 as a dependency for improved TensorFlow model export compatibility.
• Python Version Check Fix: Corrected version comparison logic for Jetson devices to prevent unnecessary dependency downgrades.
• Documentation Enhancements: Improved clarity, formatting, and examples across multiple sections, including YOLO11 integration and dataset usage.
• Customizable Text Colors: Introduced a txt_color parameter for detection result annotations, allowing users to specify RGB text colors.

🎯 Purpose & Impact

• Improved Training Flexibility: The single-class training fix ensures smoother workflows for users focusing on specific object categories.
• Enhanced Export Reliability: TensorFlow users benefit from more stable and efficient model exports, especially for edge AI applications.
• Broader Compatibility: Fixes for Python version checks enhance usability on Jetson devices, avoiding unnecessary compatibility issues.
• User-Friendly Documentation: Updated guides and examples make it easier for users to understand and utilize YOLO11 features effectively.
• Customizable Visual Outputs: The new txt_color parameter allows for more professional and adaptable visualizations, aligning with project-specific requirements. 🌈

These updates collectively enhance the usability, flexibility, and professional output of YOLO11, catering to a wide range of user needs. 🌟

What's Changed

• Add ai-edge-litert>=1.2.0 to exporter.py by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19727
• Fix Autobackend Python version check by @Auc7us in https://github.com/ultralytics/ultralytics/pull/19722
• Docs /usage updates by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19731
• Fix broken links in docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19734
• Bump astral-sh/setup-uv from 4 to 5 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/19736
• Documentation fixes and improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19749
• Expose txt_color parameter for Results plots by @zanaries in https://github.com/ultralytics/ultralytics/pull/19718
• ultralytics 8.3.92 Fix single_cls training cache error by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19739

New Contributors

• @zanaries made their first contribution in https://github.com/ultralytics/ultralytics/pull/19718
• @Auc7us made their first contribution in https://github.com/ultralytics/ultralytics/pull/19722

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.91...v8.3.92

🌟 Summary

The YOLO 8.3.91 release focuses on simplifying TensorFlow installation, enhancing export compatibility for ARM64/Linux platforms, improving dataset handling, and refining documentation and visualization features. 🚀

---

📊 Key Changes

• TensorFlow Installation Simplification: Streamlined TensorFlow setup for smoother exports, with updated dependency requirements.
• Export Enhancements:
  • Improved support for ARM64 and Linux platforms for TFLite and TensorFlow.js exports.
  • Clearer error messages for unsupported configurations.
• Dataset Handling Improvements:
  • Automatic fallback logic for missing val or test dataset splits.
  • Enhanced logging of image batches during training for better Comet integration.
• Visualization Updates:
  • Adjusted font sizes in image annotations for better readability.
• Documentation Updates:
  • Added comparisons between YOLO models (e.g., YOLO11n-seg) and Meta's SAM models, highlighting YOLO's efficiency.
  • Included new social media links (e.g., WeChat) for broader accessibility.

---

🎯 Purpose & Impact

• Improved Compatibility: Export processes are now more robust on ARM64 and Linux systems, reducing platform-specific issues. 🛠️
• Enhanced User Experience:
  • Simplified TensorFlow installation and clearer error messages save users time and effort.
  • Better dataset handling minimizes manual intervention and errors. 🗂️
• Visualization & Logging:
  • Improved image annotations and logging streamline debugging and training workflows. 📊
• Documentation Clarity:
  • Performance comparisons help users select the best model for their needs.
  • Expanded social media presence ensures easier access to resources and support. 🌐

This release ensures smoother workflows, better platform support, and improved usability for developers and users alike. 🚀

What's Changed

• Improved visualization of labels in classify task for better Comet integration by @yaricom in https://github.com/ultralytics/ultralytics/pull/19700
• Add WeChat social icon to Docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19702
• Update SAM vs YOLO benchmarks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19705
• ultralytics 8.3.91 Simplify tensorflow installation by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19712

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.90...v8.3.91

🌟 Summary

The Ultralytics v8.3.90 release introduces several key updates, including a fix for memory usage calculations on Apple MPS devices, documentation enhancements, and architectural optimizations across YOLO models. 🚀

📊 Key Changes

• MPS Memory Fix: Updated memory usage calculation for Metal Performance Shaders (MPS) devices to improve accuracy using psutil.virtual_memory().percent.
• YOLO Model Optimizations: Reduced the number of layers across YOLO11, YOLOv8, and YOLOv9 models for improved efficiency without compromising performance.
• Documentation Improvements: Enhanced clarity, formatting, and consistency across multiple files, including detailed descriptions for methods and parameters.
• Logging Enhancements: Improved logging behavior for better user control and debugging.
• C++ Example Fix: Resolved input image dimension handling in YOLOv8 C++ inference code.
• Default Solution Handling: Added fallback to "count" solution when no solution name is provided in YOLO commands.

🎯 Purpose & Impact

• Enhanced User Experience:
  • Apple MPS users benefit from more accurate memory tracking, optimizing resource management during training or inference. 🍎
  • Clearer documentation and improved logging make the platform more user-friendly and easier to debug.
• Performance Boost:
  • YOLO model optimizations lead to faster inference times and reduced computational overhead, enhancing efficiency for various tasks. ⚡
• Developer Convenience:
  • Fixes in C++ examples and default solution handling streamline workflows for developers.

This release ensures better compatibility, performance, and usability for both developers and end-users. 🌟

What's Changed

• Documentation improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19667
• Fix links.yml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19665
• Fix layer counts in model YAMLs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19663
• Fix verbose for Solutions by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19651
• Fix formatToSquare bug in YOLOv8 C++ example by @matriox1003 in https://github.com/ultralytics/ultralytics/pull/19653
• Fix solutions CI by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19675
• Large Python files documentation update by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19695
• Add uv pip install for Raspberry Pi CI Benchmarks and Tests by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/17912
• ultralytics 8.3.90 Fix MPS get_memory() error by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19686

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.89...v8.3.90

🌟 Summary

The Ultralytics 8.3.89 release focuses on improving dependency management, enhancing compatibility with NVIDIA Jetson devices, and refining documentation for better usability. 🚀

📊 Key Changes

• Dependency Management: Updated --index-strategy to unsafe-best-match for safer and more reliable package installations.
• Jetson Compatibility: Adjusted TensorFlow.js versions for Jetson JetPack 4/5 to ensure seamless compatibility.
• Documentation Updates: Standardized code examples across the repository to Python's interactive shell style (>>>) for improved clarity and consistency.
• Stale Workflow Improvements: Enhanced GitHub workflows to better manage stale pull requests and issues.
• Version Update: Incremented the Ultralytics version to 8.3.89.

🎯 Purpose & Impact

• Enhanced Reliability: The updated dependency strategy minimizes potential conflicts during package installations, ensuring smoother setups. 🛠️
• Broader Hardware Support: Improved TensorFlow.js compatibility ensures better performance on NVIDIA Jetson devices, benefiting edge AI applications. 🤖
• Developer Experience: Standardized documentation makes it easier for developers to understand and implement examples, boosting productivity. 📚
• Efficient Project Management: Improved stale workflows help maintain a cleaner and more organized repository. 🚀
• Version Alignment: The version bump reflects these enhancements, signaling progress and stability.

What's Changed

• Updated stale actions rules by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/16563
• Align code examples with Google-style by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19496
• ultralytics 8.3.89 TensorFlow 2.19.0 compatibility updates by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19668

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.88...v8.3.89

🌟 Summary

The v8.3.88 release of Ultralytics introduces significant enhancements and new features across its solutions, focusing on object detection, segmentation, privacy tools, and advanced analytics. These updates aim to provide more robust, versatile, and user-friendly tools for computer vision tasks.

📊 Key Changes

• New Solutions Added:
  • ObjectBlurrer: Blurs detected objects for privacy protection.
  • ObjectCropper: Crops and saves detected objects for further analysis.
  • InstanceSegmentation: Enables segmentation with masks and annotations.
  • VisionEye: Simulates human observation by mapping objects to a vision anchor point.
• Enhanced Analytics:
  • Improved support for various chart types (line, pie, bar, area) with better visualization and customization.
  • Unified structure for analytics results.
• Refinements in Object Tracking:
  • Improved handling of bounding boxes and tracking across frames.
  • Enhanced region-based counting and queue management.
• Bug Fixes:
  • Addressed bounding box offset issues in YOLOv8 C++ inference, ensuring accurate object detection and scaling.

🎯 Purpose & Impact

• 🛡️ Privacy Protection: The ObjectBlurrer ensures sensitive information in images or videos is obscured, making it ideal for security and compliance use cases.
• ✂️ Data Preparation: ObjectCropper simplifies dataset creation by extracting and saving detected objects.
• 🎨 Advanced Visualization: InstanceSegmentation and VisionEye provide detailed insights into object relationships and spatial mapping, enhancing analytical capabilities.
• 📈 Improved Analytics: The updated analytics tools allow users to derive actionable insights from data with visually appealing and informative charts.
• 🚀 Enhanced Accuracy: Fixes in bounding box handling improve detection precision, benefiting applications like autonomous systems and surveillance.

These updates make Ultralytics solutions more powerful and adaptable, catering to a wide range of industries and use cases.

What's Changed

• Fix detection box offset bug in YOLOv8 example model inference results by @matriox1003 in https://github.com/ultralytics/ultralytics/pull/19639
• Docs: Update Banner by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/19643
• Improved Examples documentation by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19645
• ultralytics 8.3.88 Solutions refactor and improvements by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18491

New Contributors

• @matriox1003 made their first contribution in https://github.com/ultralytics/ultralytics/pull/19639

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.87...v8.3.88

🌟 Summary

The v8.3.87 release introduces a new Results.to_html() method for exporting inference results in HTML format, alongside various usability, compatibility, and documentation improvements.

📊 Key Changes

• HTML Export for Results: Added to_html() method to convert detection results into a web-friendly HTML format.
• Documentation Enhancements: Improved clarity and consistency in method descriptions and added a dedicated page for the COCO128 dataset.
• Compatibility Updates:
  • Added support for Ubuntu ARM CI runners and removed QEMU emulator for ARM64 Docker builds.
  • Restricted OpenVINO dependency to avoid version conflicts.
• Memory Management: Optimized GPU memory clearing to trigger only when usage exceeds 90%.
• Bug Fixes:
  • Resolved bounding box out-of-bounds issues in MNN examples.
  • Fixed multi-stream video result saving to avoid overwriting files.
• Classification Model Updates: Adjusted FLOPs calculations for classification models to align with the default image size of 224 pixels.
• Comet Integration: Added support for logging segmentation annotations in Comet.

🎯 Purpose & Impact

• Improved Usability:
  • The to_html() method simplifies sharing and visualizing inference results in a browser-friendly format.
  • Enhanced documentation, including COCO128 details, aids developers in testing and debugging.
• Broader Compatibility:
  • ARM support ensures better performance and inclusivity for ARM-based systems.
  • OpenVINO version constraints prevent compatibility issues.
• Optimized Performance:
  • Smarter GPU memory management avoids unnecessary slowdowns during training.
  • Faster ARM64 Docker builds reduce build times from 8 minutes to 2 minutes.
• Enhanced Functionality:
  • Segmentation support in Comet improves tracking and visualization for segmentation tasks.
  • Bug fixes ensure more reliable and accurate results across various use cases.

What's Changed

• Enable Ubuntu ARM GitHub CI runners by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19596
• Add COCO128 dataset page to docs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19594
• Fix BBox Out-of-Bounds Issue in MNN Example by @jules-ai in https://github.com/ultralytics/ultralytics/pull/19597
• Add extra step for installing tkinter for annotator tool by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19534
• Update classification model FLOPs with imgsz=224 by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19593
• Exclude ort.pyke.io from link checking by @decahedron1 in https://github.com/ultralytics/ultralytics/pull/19554
• Solutions init conditional import by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19589
• Constrain OpenVINO versions to >=2024.0.0,!=2025.0.0 by @aleksandr-mokrov in https://github.com/ultralytics/ultralytics/pull/19481
• Fix result saving issue when multi-stream detection by @XevenQC in https://github.com/ultralytics/ultralytics/pull/19297
• Remove QEMU emulator in favor of Github-hosted Arm64 Ubuntu runners by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19604
• Add Segmentation support in Comet logging by @yaricom in https://github.com/ultralytics/ultralytics/pull/19618
• Revert shell=True for model.tune() by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19625
• Remove parking slots annotator from tests by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19624
• Update docs.yml Ruff fixes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19626
• empty_cache() only if GPU utilization above 90% by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/18949
• ultralytics 8.3.87 New Results.to_html method for inference outputs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19161

New Contributors

• @decahedron1 made their first contribution in https://github.com/ultralytics/ultralytics/pull/19554
• @aleksandr-mokrov made their first contribution in https://github.com/ultralytics/ultralytics/pull/19481
• @XevenQC made their first contribution in https://github.com/ultralytics/ultralytics/pull/19297

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.86...v8.3.87

🌟 Summary

The v8.3.86 release focuses on improving dataset handling, enhancing code consistency, and fixing minor issues to streamline workflows for users working with various datasets and models.

📊 Key Changes

• Dataset YAML Refactor: Improved dataset configuration files for better readability and functionality.
  • Enhanced autodownload and conversion scripts for datasets like COCO, VOC, ImageNet, and more.
  • Unified string formatting to double quotes (") across YAML files.
  • Added detailed comments and docstrings for clarity.
• UTF-8 Encoding Compliance: Updated all file operations to explicitly use UTF-8 encoding for better cross-platform compatibility.
• Keypoint Loss Fix: Adjusted keypoint loss calculations to ensure only relevant dimensions are divided by stride tensors.
• Documentation Updates:
  • Fixed example code in SAM 2 model documentation.
  • Embedded updated YouTube tutorials for YOLO11 training and batch inference.
• Code Cleanup: Removed redundant imports and improved path handling using modern Python practices.

🎯 Purpose & Impact

• Enhanced Dataset Handling:
  • Simplifies dataset preparation with improved autodownload and conversion scripts. 🛠️
  • Reduces errors with better file handling and clear documentation. 📚
• Cross-Platform Compatibility:
  • Ensures consistent behavior across operating systems by enforcing UTF-8 encoding. 🌍
• Improved Model Accuracy:
  • Fixes in keypoint loss calculations enhance precision in keypoint-based tasks. 🎯
• User Experience:
  • Updated tutorials and documentation make it easier for users to train and deploy models effectively. 🎥
• Code Maintainability:
  • Cleaner and more consistent codebase improves readability and reduces potential bugs. 🧹

This release is a quality-of-life update that enhances both backend functionality and user-facing resources, making it easier for developers and researchers to work with Ultralytics tools. 🚀

What's Changed

• Update sam-2.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19573
• Remove superfluous LOGGER import from data/converter by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/19577
• Add https://youtu.be/ZN3nRZT7b24 to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19565
• Exclude keypoint visibility from stride division by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19567
• Remove extra import from cfg/__init__ by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/19576
• Adds open encoding per PEP-597 by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/19578
• Set dynamic from metadata by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19560
• ultralytics 8.3.86 Refactor dataset YAML autodownload scripts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19579

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.85...v8.3.86

🌟 Summary

This release, v8.3.85, focuses on improving TensorRT export functionality and refining ONNX segmentation examples for better performance and usability. 🚀

---

📊 Key Changes

TensorRT Updates (Main Focus of Release):

• Fixed max_shape Calculation Bug: Resolved inconsistent calculations during TensorRT export with non-zero workspace values.
• Improved Default Behavior: Ensured workspace settings default to 0 when not specified.

ONNX Segmentation Example Enhancements:

• Streamlined Preprocessing/Postprocessing: Simplified the example to use a more efficient and flexible implementation, including key parameters like iou, imgsz, and conf.
• Unified Confidence Threshold: Updated ONNX example to align with YOLO conventions for clarity (conf, iou).
• Optimized Mask Handling: Enhanced segmentation mask functionality for better accuracy and resource efficiency.
• Adaptive Backend Setup: Integrated GPU support automatically when available, improving ease of use.

---

🎯 Purpose & Impact

• For TensorRT Users:

  • Purpose: Addressed critical bugs in TensorRT export needed for high-performance inference engines.
  • Impact: Accurate shape calculation ensures smoother deployments of dynamic models with TensorRT, preventing crashes or incorrect behavior. Users exporting YOLO models into .engine format, especially with non-zero workspaces, will experience stable exports.

• For ONNX Developers:

  • Purpose: Simplify, optimize, and enhance usability of segmentation examples using ONNX Runtime.
  • Impact: Developers using ONNX Runtime benefit from faster setup, more intuitive parameter usage, and reliable mask-based object segmentation results.

• General Improvements:

  • These updates improve the overall developer experience and model deployment across various inference backends, enabling seamless workflows for cutting-edge scenarios.

---

This update primarily strengthens export and inference capabilities 🎯, while making the process more robust for advanced and typical users alike.

What's Changed

• Cleanup and fix ONNX segment example by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19551
• ultralytics 8.3.85 TensoRT export max_shape fix by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19541

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.84...v8.3.85

🌟 Summary

The v8.3.84 release brings improvements to YOLO's segmentation handling, documentation clarity, and usability, with a focus on filtering invalid outputs and refining user guidance. 🎉

📊 Key Changes

• 🚀 Segmentation Optimization: YOLO now filters out predictions with empty masks, resulting in more accurate and valuable outputs.
• 📚 Updated Documentation Features:
  • Added a correct YouTube link for SAHI Tiled Inference for better instructional resources. 🎥
  • Improved code examples for consistent and clear understanding of critical tools like Colors class and merge_equals_args. ✨
• ⚙️ Validation Enhancements: Restricted save_hybrid mode to only detection tasks, preventing incorrect usage and evaluation inaccuracies.

🎯 Purpose & Impact

• 🧹 Cleaner Segmentations: By removing predictions lacking usable masks, segmentation results are more reliable and efficient. This streamlines workflows and avoids irrelevant data.
• ✅ Easier Learning & Usage: Updated documentation improves the user experience through better resource links and clearer code examples, easing the learning curve for developers and users.
• ⚠️ Prevents Missteps: Limiting save_hybrid ensures omission of scenarios that could lead to misinterpretation of validation outputs, particularly for non-detection models.

This update is all about boosting the quality and usability of YOLO tools, paving the way for more productive and error-free model usage! 🏆

What's Changed

• Add https://youtu.be/ILqMBah5ZvI to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19532
• Fix reference docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19528
• Disable save_hybrid for OBB and update docs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19531
• ultralytics 8.3.84 Remove predictions with no masks by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19537

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.83...v8.3.84

🌟 Summary

Ultralytics release v8.3.83 focuses on refining image augmentation to make color transformations more natural while clarifying validation parameter requirements in the documentation. 🎉

---

📊 Key Changes

• Image Augmentation Adjustments:

  • Reverted hue, saturation, and value (HSV) augmentation logic back to a relative shift approach for more natural-looking image transformations.
  • Fixed an issue with hue adjustments to align with the original implementation for improved consistency.
  • Ensured constraints, like maintaining pure white to avoid unnatural color changes, remain enforced. 🎨

• Documentation Improvement:

  • Clarified the batch parameter for validation, specifying that it must be a positive integer to avoid confusion about unsupported functionality like AutoBatch in validation. 📚

---

🎯 Purpose & Impact

• For Developers and Users:
  • 🎨 Enhanced Visual Realism: The relative shift method restores the natural look of augmented images, which is important for data augmentation and model training. This could improve model performance on tasks requiring realistic visual data.
  • 🛠 Accurate Transformations: Bug fixes ensure reliable and consistent adjustments during augmentation, preventing potential training issues.
  • 📚 Eased Configuration: Improved documentation reduces ambiguity in setting validation parameters, simplifying usage for both new and experienced users.

This release is a step forward in delivering both technical precision and user efficiency! 🚀

What's Changed

• Fix batch description for validation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19504
• ultralytics 8.3.83 Revert saturation and value augmentation to relative shift by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19515

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.82...v8.3.83

🌟 Summary

Ultralytics v8.3.82 enhances ONNX model export functionality, improves preprocessing for ONNXRuntime examples, and streamlines compatibility across various hardware setups. 🚀

📊 Key Changes

• ONNX FP16 Export Fix: Introduced a workaround (arange_patch) for exporting ONNX models with both dynamic and half options, avoiding incompatibilities in the PyTorch torch.arange function.
• Preprocessing Accuracy Boost: Improved image handling (aspect ratio, resizing, and padding) for ONNXRuntime examples to ensure more precise object detection.
• Broader MNN Test Coverage: Enabled MNN export testing on Raspberry Pi to ensure compatibility.
• Dataset Configuration Streamlining: Refined open-images-v7.yaml to centralize dataset directory management for clarity and maintainability.

🎯 Purpose & Impact

• ONNX Export Enhancements:

  • Purpose: Solve PyTorch incompatibility issues during ONNX model export with FP16 precision in dynamic mode.
  • Impact: Developers can now seamlessly export high-performance ONNX models while avoiding device-specific restrictions, improving workflows on GPUs.

• Accurate Preprocessing:

  • Purpose: Fix preprocessing bugs in ONNXRuntime inference to ensure image resizing, aspect ratio, and padding align with expected outputs.
  • Impact: Results in more consistent and reliable object detection performance between PyTorch and ONNX models.

• Compatibility Extension:

  • Purpose: Enable MNN export tests on Raspberry Pi to ensure cross-platform functionality.
  • Impact: Broader hardware compatibility for MNN format users, improving reliability on low-cost systems like Raspberry Pi.

• Dataset Management Improvements:

  • Purpose: Streamline configuration for the Open Images V7 dataset.
  • Impact: Easier and more consistent dataset setup, particularly useful for developers managing large datasets.

This update improves model export workflows, ensures consistent inference results across platforms, and expands compatibility for developers using diverse environments. 🌐

What's Changed

• Enable mnn in Raspberry Pi Tests by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19483
• Fix ONNX Example letterboxing by @quangdungluong in https://github.com/ultralytics/ultralytics/pull/19486
• Fix dataset_dir error with open-images-v7.yaml by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19475
• ultralytics 8.3.82 ONNX dynamic and half export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19464

New Contributors

• @quangdungluong made their first contribution in https://github.com/ultralytics/ultralytics/pull/19486

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.81...v8.3.82

🌟 Summary

The Ultralytics v8.3.81 release resolves a critical memory management issue in model validation while introducing key updates to documentation, testing workflows, and system reporting for enhanced usability and efficiency. 🚀✨

---

📊 Key Changes

• 🧹 Memory Leak Fix in Validation Modules:
  Resolved circular references in metrics (on_plot) across validators (DetectionValidator, PoseValidator, etc.) to avoid CPU memory leaks during repeated evaluations.
• 📚 Documentation Enhancements:
  Added examples for annotators in the library, and fixed/corrected metadata instructions for Triton guides and links in SAM 2 documentation.
• 🐧 Raspberry Pi CI Improvements:
  Re-enabled Raspberry Pi CI workflow with improved benchmarks and added cleanup for self-hosted CI runners.
• 🔍 Installation Diagnostics Update:
  Added project root installation path in the system environment summary for easier debugging.
• 📊 Better Table Sorting in Docs:
  Improved table sorting functionality for file sizes, numbers, and dot-separated values.

---

🎯 Purpose & Impact

• Memory Management Improvement:
  Eliminates residual memory accumulation when evaluating models repeatedly in scripts, preventing CPU overloading and Out-Of-Memory (OOM) errors. This fix improves the efficiency and reliability of long-running evaluation pipelines. 🧠🔥
• User-Friendly Documentation:
  New usage examples and enhanced instructions make it easier for developers to understand and implement key features. Correcting SAM 2 links ensures access to the right research resources. 📚💡
• Expanded Testing Coverage:
  Re-enabling Raspberry Pi CI and adding benchmarks enhances testing on diverse hardware, ensuring broader reliability across platforms. Improved CI cleanup reduces risks of resource conflicts. 🌐🔧
• Debugging Simplification:
  Including installation paths in diagnostics facilitates resolving Python environment-related issues for developers. 🔍🛠️
• Enhanced Sorting for Visual Data:
  Improved table sorting in docs boosts usability, especially when navigating large datasets or results with complex numbering or file size values. 📈✨

Overall, this update strengthens performance, stability, and developer experience, while prioritizing clarity and usability for the broader community. 🎉

What's Changed

• Update Raspberry Pi 5 CIs with 16GB model by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19306
• New: add SAM and SAM-2 notebook in docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19461
• Add installation path to yolo checks output by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19463
• Clarify metadata instructions in Triton docs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19457
• Fix Docs table number sorting by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19455
• Add examples for plotting reference by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19439
• Update SAM 2 research paper links by @joshua-dean in https://github.com/ultralytics/ultralytics/pull/19465
• Clean up Raspberry Pi CI runners by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19478
• ultralytics 8.3.81 Fix Metrics on_plot circular references by @RemiPT in https://github.com/ultralytics/ultralytics/pull/19318

New Contributors

• @joshua-dean made their first contribution in https://github.com/ultralytics/ultralytics/pull/19465
• @RemiPT made their first contribution in https://github.com/ultralytics/ultralytics/pull/19318

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.80...v8.3.81

🌟 Summary (single-line synopsis)

Version 8.3.80 features crucial upgrades to YOLO-NAS handling, improved configuration merging for smoother exports, enhanced documentation interactivity, and compatibility refinements for key frameworks. 🚀

---

📊 Key Changes (bullet points highlighting any major changes)

• 🔄 Model Export Enhancements: Fixed YOLO-NAS export by integrating default configs (DEFAULT_CFG_DICT) into model attributes for higher flexibility.
• 🧠 RBOX Regularization: Refined bounding box angle calculations for consistency in the OpenCV definition.
• 📋 Documentation Upgrades: Introduced sortable tables for performance data to improve clarity and usability across pages.
• 🔧 Framework Compatibility: Constrained OpenVINO dependency to version >=2024.0.0,<2025.0.0 and updated outdated function calls for compatibility.
• 🐳 Docker Workflow Fix: Removed deprecated numpy dependency to resolve CI errors and streamlined build conditions.

---

🎯 Purpose & Impact (bullet points explaining any benefits and potential impact to users)

• 🌟 Improved Export Reliability: The configuration merging for YOLO-NAS ensures a more robust process, minimizing errors during model deployment and export.
• 🧮 Better Prediction Accuracy: Regularizing rboxes improves detections by addressing angle inconsistencies in bounding boxes.
• 🖱️ Enhanced User Experience: Sortable tables in documentation allow users to interactively organize performance metrics, making comparisons easier.
• ✅ Smoother OpenVINO Exports: Version constraints and function updates enhance stability and prepare the platform for future releases without breaking changes.
• 🔄 Cleaner CI Pipelines: Removing deprecated dependencies ensures more reliable builds, benefitting developers contributing to the codebase.

This update underscores Ultralytics' focus on improving usability, compatibility, and export-related functionalities for smoother workflows and precise model performance. 🚀

What's Changed

• Optimize Sony IMX500 doc by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19421
• Constrain OpenVINO versions to >=2024.0.0,<2025.0.0 by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/19122
• New Docs sortable tables by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19376
• Fixing tflite export CI error by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/19422
• Fix rbox regularization by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19429
• ultralytics 8.3.80 Fix YOLO-NAS export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19426

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.79...v8.3.80

🌟 Summary

The v8.3.79 release of the Ultralytics YOLO framework introduces crucial bug fixes, enhancements to performance, and documentation updates. The primary focus is on correcting HSV augmentation mechanics and refining various code components for better reliability and usability. 🛠️✨

---

📊 Key Changes

• HSV Augmentation Fix: Resolved incorrect hue, saturation, and value shifts during data augmentation, ensuring consistent color transformations. 🎨
• YOLO12 Benchmark Refinement: Fixed performance metrics for YOLO12 models in documentation, updating speed and mAP comparisons ensuring accurate representations. 📈
• Docker Streamlining: Removed redundant tensorrt-cu12 dependency and added environment checks for better CI validation and Docker compatibility. 🐳
• Comet Integration Fix: Corrected class mapping index misalignment in Comet logging to avoid errors and misinterpretations in visualized data. 📋
• Single-Class Model Consistency: Enforced single-class model output alignment by adjusting nc and names parameters during training. 🚀
• Documentation Updates: Improved accuracy in multiple areas, embedding a YouTube interactive guide for YOLO12, and updating outdated citation links for research credibility. 📚
• General Code Cleanup: Enhanced readability by adhering to PEP8 conventions (snake_case), simplifying code for maintainability. 🖋️

---

🎯 Purpose & Impact

• Improved Data Augmentation: Provides users with accurate HSV augmentations, leading to better model robustness and performance during training.
• Benchmark Accuracy: Ensures users receive correct performance metrics allowing better model selection and easier comparison across versions.
• Enhanced Docker Usability: Smoother image builds and reduced dependency issues result in a simplified development experience. Perfect for CI pipelines!
• Better Logging: Developers benefit from accurate class mappings in Comet analytics, avoiding confusion when examining predictions.
• Single-Class Models Alignment: Offers consistency in model behavior, improving usability for tasks like binary classification.
• Improved Documentation: Users gain easier access to tutorials, accurate technical references, and a better learning experience overall.
• Cleaner Codebase: Enhances long-term maintainability and aligns code with modern Python standards, beneficial for both contributors and end-users.

This release makes strides in improving accuracy, user experience, and usability for developers and researchers alike. 🚀🌟

What's Changed

• Fix YOLO12 cfg links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19332
• Fix mkdocs_github_authors.yaml author LexBarou by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19334
• Fix keypoints data display in pose.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19337
• Fix YOLO12 performance table by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19333
• Fix benchmark.js by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19351
• Fix TensorRT export issue by removing redundant package by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19343
• Add https://youtu.be/SMSJvjUG1ko to yolo12.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19358
• Update model_data.py authors by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19365
• Assign variable dt before reference in exception block by @sjhpark in https://github.com/ultralytics/ultralytics/pull/19349
• Use chart.js@latest by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19372
• YOLOv3 paper URL update by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19378
• Add imx export tip by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19404
• Det models with nc=1 when single_cls=True by @kevinconka in https://github.com/ultralytics/ultralytics/pull/19381
• Fixed Comet integration to use class map aligned index when trying to get class name by @yaricom in https://github.com/ultralytics/ultralytics/pull/19408
• Simplify auto_annotate function by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/19400
• Updates imgToAnns variable to snake_case by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/19402
• Fix Grad strides do not match bucket view strides warning for YOLO12 DDP training by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19395
• Updates scaleFill argument to snake_case by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/19401
• Check environment in docker.yml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19418
• Disable Dockerfile tests temporarily by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19420
• Fix incorrect metrics when nothing is detected by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19362
• ultralytics 8.3.79 Fix shift in HSV augmentation by @picsalex in https://github.com/ultralytics/ultralytics/pull/19311

New Contributors

• @picsalex made their first contribution in https://github.com/ultralytics/ultralytics/pull/19311
• @kevinconka made their first contribution in https://github.com/ultralytics/ultralytics/pull/19381
• @sjhpark made their first contribution in https://github.com/ultralytics/ultralytics/pull/19349

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.78...v8.3.79

🌟 Summary

This release, v8.3.78, brings an exciting new model to the family: YOLO12 🚀, featuring an attention-centric design for superior accuracy and efficiency across a variety of computer vision tasks.

---

📊 Key Changes

• Introduction of YOLO12 Models:

  • New Architecture: Incorporates attention mechanisms like Area Attention, R-ELAN, and FlashAttention for optimized performance.
  • Comprehensive Task Support:
    • Object detection, segmentation, pose estimation, classification, and oriented bounding box (OBB) detection.
  • Benchmark Improvements:
    • Demonstrates higher mAP (mean Average Precision) and efficiency compared to YOLO10/YOLO11 and competitors like RT-DETR.

• Model-Specific Enhancements:

  • Introduced multiple YOLO12 variants (n, s, m, l, x) catering to different computing environments such as cloud systems and edge devices.
  • Added new task-focused configurations for image classification, pose estimation, and segmentation.

• Documentation Updates:

  • YOLO12 now included in detailed model documentation with performance metrics and usage examples.
  • Extensive references, including benchmarks for comparison with leading global detection models.

• Code Simplifications and Bug Fixes:

  • ONNX Run-Time Fixes: Improved device handling and tensor reshaping for ONNX users.
  • TFLite Export Cleanup: Removed redundant parameters for simpler TensorFlow Lite export logic.
  • Code Refinement: Enhanced readability and maintainability across inference and export pipelines.

---

🎯 Purpose & Impact

• Purpose:

  • YOLO12 brings a paradigm shift in accuracy and efficiency by adopting attention mechanisms tailored for real-time object detection.
  • Streamlines codebase for easier maintenance and integration in diverse projects.

• Impact:

  • Developers gain access to cutting-edge state-of-the-art models excelling in versatility, speed, and precision.
  • Tasks like multi-object detection, segmentation, and pose estimation become more accessible for smaller devices (e.g., edge devices).
  • Improved user experience with easier model selection, robust export support, and refined prediction outputs.

🔮 This update is not only a leap forward in technological advancement but also a commitment to making intelligent vision accessible to all.

What's Changed

• Remove unused parameters from export_tflite by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19319
• Fix ONNX RuntimeError with dynamic WorldModel by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19322
• Add YOLO12 model info by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19328
• Add https://youtu.be/BPYkGt3odNk to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19331
• Refactor with simplifications by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19329
• update model_data.py by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19330
• ultralytics 8.3.78 new YOLO12 models by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19325

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.77...v8.3.78

🌟 Summary

The YOLOv8 v8.3.77 release introduces a significant performance optimization for ONNX Runtime segmentation models, as well as essential compatibility enhancements and minor fixes. 🚀

---

📊 Key Changes

• 🚀 YOLOv8-Segment Optimization (ONNX Runtime)

  • Post-processing was drastically improved for both GPU (91.7% faster) and CPU (58.9% faster), resulting in a total inference speedup of up to 32.4%.

• 🔧 Optional thop Dependency Support

  • Made the thop library optional by handling its absence gracefully to avoid errors in environments lacking it.

• 📂 ONNX Export Improvements

  • Streamlined ONNX dynamic model export logic by revising type handling, ensuring better reliability and maintainability.

---

🎯 Purpose & Impact

• Faster and More Efficient Inference ⚡

  • The optimization of YOLOv8-Segment ONNX Runtime reduces latency significantly, improving user experiences in deployment scenarios where real-time performance is essential.

• Increased Compatibility with Minimal Setups 🤝

  • By making thop optional, the release ensures broader support for systems, including lightweight environments like Conda setups, enhancing user flexibility.

• Improved Developer Experience 🛠️

  • Simplified export logic and code maintainability ensures a more robust development and debugging process for ONNX users.

These updates collectively enhance usability, efficiency, and reliability for YOLOv8 users across diverse applications. 🎉

What's Changed

• Fix dynamic export with YOLO World by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19304
• Revert "Fix dynamic export with YOLO World" by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19308
• Allow missing thop package by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19314
• ultralytics 8.3.77 faster YOLOv8-Segment ONNX Runtime example by @AdnanEkici in https://github.com/ultralytics/ultralytics/pull/19312

New Contributors

• @AdnanEkici made their first contribution in https://github.com/ultralytics/ultralytics/pull/19312

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.76...v8.3.77

🌟 Summary

The release of Ultralytics v8.3.76 introduces improved dynamic batch inference for ONNX models with NMS export, a better object tracking experience, and various code and documentation enhancements. 🚀

---

📊 Key Changes

• Dynamic Batch Improvements:

  • Solved issues with dynamic=True and nms=True where the batch size was fixed at export.
  • Enabled dynamic input handling by introducing padding for unmatched batch sizes during ONNX exports.

• Tracking Enhancements:

  • Addressed errors when processing Torch tensors in model.track().
  • Improved integration of original input images with trackers for accuracy.

• Performance Accuracy:

  • Fixed GPU memory conversion errors for logging VRAM usage to ensure accurate reporting.

• Documentation Updates:

  • Standardized formatting in documentation for better consistency.
  • Enhanced examples for interpreting prediction results across detection, pose, segmentation, and more.

• Other Code Refinements:

  • Addressed layer miscount in logging by ensuring layers with no parameters are also displayed.
  • Enhanced GitHub issue templates for clearer differentiation between bug reports and feature requests.

---

🎯 Purpose & Impact

• 🛠 Improved Model Deployment: Dynamic padding during export ensures robust handling of varying batch sizes while maintaining compatibility with ONNX workflows.
• 🎥 Better Tracking: Smoother operation for streamed data and enhanced consistency in object tracking results benefit both developers and end-users.
• 📋 Accurate Logging: Correct VRAM usage metrics improve debugging and resource optimization.
• 📚 Developer Friendliness: Updated docs and examples make it easier for users to understand and utilize prediction results effectively.
• 🚀 Efficiency Boost: Code tweaks and fixes culminate in faster, more accurate model handling without disruptions.

This release addresses several community-reported issues, focusing on operational accuracy and usability across export, tracking, and development workflows! 🙌

What's Changed

• Initialize model_name attribute by @LoveAndHope-dev in https://github.com/ultralytics/ultralytics/pull/19224
• Update results.boxes docs: boxes.id may be None by @shankangke in https://github.com/ultralytics/ultralytics/pull/19227
• Fix bytes to GB conversion for logged VRAM usage by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19254
• Fix memory conversion by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19256
• Use new "Issue Type" in templates by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19257
• Add https://youtu.be/im7xBCnPURg to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19265
• Support CoreML NMS export for Segment, Pose and OBB by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19173
• Fix layer count; show layers with no params in detailed log by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19202
• Standardize str formatting in docs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19276
• Add detailed usage with demos to reCamera doc by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19275
• Add examples showing how to use result for all tasks by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19282
• Update model_data.py by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19267
• Revert "Support CoreML NMS export for Segment, Pose and OBB" by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/19273
• Fix error with torch tensor input in model.track() by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19278
• ultralytics 8.3.76 fix dynamic batch inference with NMS export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19249

New Contributors

• @LoveAndHope-dev made their first contribution in https://github.com/ultralytics/ultralytics/pull/19224

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.75...v8.3.76

🌟 Summary

The v8.3.75 release includes robust updates for improved model export compatibility, user experience, and error handling across platforms, alongside enhanced documentation and integration refinements. 🚀

---

📊 Key Changes

• Enhanced CometML Integration:

  • Transitioned to the new comet_ml.start() API for smoother experiment handling.
  • Deprecated the COMET_MODE variable, introducing COMET_START_ONLINE for consistency.

• Export Function Updates:

  • Protobuf Dependency: Added support for protobuf>=5 for TensorFlow and TFLite exports, resolving compatibility issues.
  • Edge TPU and TF.js: Addressed platform-specific limitations for ARM64 and Linux exports to prevent unsupported configuration errors.

• Documentation Improvements:

  • Updated SAM auto-annotation, YOLOv8, and export format descriptions for clarity.
  • Redesigned inference examples to use accessible publicly hosted image URLs.

• New CLI Solutions:

  • Introduced practical use cases, including object counting, workout monitoring, queue analysis, and browser-based inference with Streamlit.

• Benchmarking Added:

  • Include new comparative performance metrics for popular object detection models like Gold-YOLO, YOLO-NAS, RTDETRv3, etc.

• Windows-Specific Fix:

  • Resolved an async file write bug to improve caching reliability on Windows systems.

• Improved Timing Precision:

  • Switched to time.perf_counter() for latency measurements, ensuring greater precision during benchmarking.

---

🎯 Purpose & Impact

• Improved Experiment Tracking:

  • Seamless CometML integration provides better environment consistency and logging during training processes.

• Enhanced Export Reliability:

  • Future-proofs TensorFlow and TFLite workflows while providing early error detection for ARM64/Linux users.

• Streamlined User Experience:

  • Updated documentation and example consistency ensure clarity, especially for beginners, minimizing friction during model setup and usage.

• Greater Platform Support:

  • Addressed critical Windows and platform-specific export edge cases, enhancing cross-platform usability.

• Better Model Insights:

  • Added benchmarks empower users to make informed decisions about which object detection models to implement based on accuracy, speed, and computational cost.

This release focuses heavily on improving reliability, usability, and documentation quality while resolving critical bugs and adding more tools for diverse real-world applications.

What's Changed

• Auto-annotate and SAM docs update by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19156
• Switch to perf_counter() for latency measurement by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19177
• Force protobuf>=5 for SavedModel export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19206
• Fix Docker QEMU issues while building JetPack 6 Dockerfile by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19216
• Fix bus.jpg path in predict.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19203
• Update NMS description on export-args.md by @Buligon in https://github.com/ultralytics/ultralytics/pull/19211
• Add NMS related args to export-table.md by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19215
• Add Docs models benchmark by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/19176
• Tasks Docs updates by @LexBarou in https://github.com/ultralytics/ultralytics/pull/19181
• YOLOv8 Docs updates by @LexBarou in https://github.com/ultralytics/ultralytics/pull/19182
• Add Solutions CLI usage in quickstart.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/19160
• Fix edgetpu and tfjs exports for arm64 Linux by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/19154
• Fix windows async np.save bug by @eric80739 in https://github.com/ultralytics/ultralytics/pull/19218
• Fix print() for ConfusionMatrix for Classify task by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/19169
• Fix updating of best epoch during early stopping by @vfcosta in https://github.com/ultralytics/ultralytics/pull/19164
• ultralytics 8.3.75 Comet update to new comet_ml.start() API by @yaricom in https://github.com/ultralytics/ultralytics/pull/19187

New Contributors

• @vfcosta made their first contribution in https://github.com/ultralytics/ultralytics/pull/19164
• @eric80739 made their first contribution in https://github.com/ultralytics/ultralytics/pull/19218

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.74...v8.3.75