animal_detection

Default precision is now Precision.fp32 instead of fp16. This changes numeric output. flutter_litert 3.8.0 changed its own default for the same reason: across 29 published detection models measured on five GPUs, fp16 matched a plain-CPU reference for only about a fifth of them, while fp32 matched every model that compiled. These graphs emit pixel-space coordinates and landmark positions, and fp16 carries about three decimal digits of mantissa, so the error lands directly on output geometry. The cost is real and worth stating plainly: fp32 is a median 29.9% slower on GPU across those five GPUs, with Apple M4 the lone exception at 6.5% faster. Pass precision: Precision.fp16 explicitly to restore the previous behaviour, ideally per model and validated on your target GPU.

Pin flutter_litert to ^3.8.0. Adds an optional LiteRT Next CompiledModel backend, off by default, and pins flutter_litert to ^3.8.0.

New useCompiledModel on initializeFromBuffers. When true, every stage is initialized onto CompiledModel instead of Interpreter. Off by default, matching the same opt-in in face_detection_tflite, pose_detection and hand_detection. compiledForceCpu is accepted alongside it.

The CompiledModel backend is CPU-only, deliberately. Each stage requests {Accelerator.cpu} rather than the permissive {gpu, cpu} set. This is a correctness requirement, not a performance preference: LiteRT miscomputes two of this package's own models once the GPU accelerator is in the set, by 42.3% of the output range for the species classifier and 53.8% for the pose model, while still reporting success. Do not widen the accelerator set without measuring against a bare-CPU reference first; flutter_litert 3.7.0 ships verifyCompiledModel for exactly that check.

Whether CompiledModel is faster is per-platform and per-model. Its CPU accelerator beats the Interpreter's XNNPACK path on Apple Silicon macOS but is roughly 2x slower on iOS, so measure before enabling it.

SSD output shapes are now derived rather than assumed, so the body detector works under both backends.

Requires flutter_litert ^3.8.0, which fixes a 3x macOS CPU slowdown affecting every stage of this pipeline (ruy multithreading was inert in the previously bundled macOS dylib).

Replace the boxed nested input and output tensors in every model class with reused flat Float32Lists handed to TFLite as ByteBuffers, matching the approach in face_detection_tflite, pose_detection and hand_detection. Measured end to end on the cat_detection pipeline over a 3264x2448 photo in profile mode with PerformanceMode.auto, the full pipeline drops from 438 ms/frame to 109 ms/frame and poseOnly from about 44 ms to 15 ms. Model outputs are unchanged.

The two new helpers ImageUtils.matToFloat32Simd and matToFloat32ImageNetSimd use OpenCV's vectorized path and accept an optional caller buffer. Both are asserted equal to the per-pixel loops they replace: worst deviation 5.96e-8 (one float32 ULP) for the plain path, 7.15e-7 for the ImageNet affine. ImageUtils.matToFloat32 and matToFloat32ImageNet are retained.

Fix ImageUtils.cropAndResize returning CropMetadata built from pre-truncation floats while cropping an integral region. Callers mapped normalized coordinates against an origin up to 1px from where the crop actually began, and against a slightly too-large extent. Measured over the 311-image CatFLW holdout with real localizer boxes, this placed landmarks +0.61px right and +0.52px down of ground truth and cost 0.255 NME_IOD, rising to 1.14 at the 95th percentile, with 72% of images improved by the fix. The Python training pipelines normalize against the integer crop, so this also aligns inference with how the models were trained.

This changes returned landmark coordinates. Downstream packages should bump their own pipeline version so cached detections re-evaluate.

Fix ModelDownloader.modelHrnet, which requested superanimal_hrnet_w32_256_float16.tflite while the release publishes superanimal_hrnet_w32_float16.tflite. AnimalPoseModel.hrnet therefore failed with an HTTP 404 on first use and had never worked. The private filename constant now derives from the public one so the two cannot drift.

Reject a configured input size that disagrees with the bundled model, via the new internal assertSquareInputSize. Interpreter.resizeInputTensor accepts a shape the model was not trained for without reporting an error, and inference then returns finite but meaningless values: feeding a 256px landmark model at 384px measured NME_IOD 67.3 against a correct 3.5 while every output stayed in range. All six model classes now validate against getInputTensor(0).shape at initialization.

Stop tracking .DS_Store files, which were included in the published archive.