dog_detection 4.1.1

Dog face and landmark detection using on-device LiteRT (formerly TensorFlow Lite) models.

4.1.1

• detectFromCameraImage now reads a desktop frame's byte order from CameraImage.format.raw instead of assuming BGRA only on macOS, so camera streams decode correctly on camera_desktop 2.0.0 (BGRA on every desktop platform) as well as 1.x (RGBA on Linux and Windows). Linux and Windows frames from camera_desktop 2.0.0 were previously decoded with red and blue swapped. Other format.raw values keep the old default, and an explicit isBgra still wins, so drop any isBgra: Platform.isMacOS you pass.
• Depend on animal_detection ^4.1.2 and flutter_litert ^3.9.2, which provides the frame-based default.
• Correct the useCompiledModel docs: it has defaulted to on since 3.0.1, not off.
• The example app depends on camera_desktop ^2.0.0.
• No API changes.

4.1.0

• Depend on animal_detection ^4.1.1 and flutter_litert ^3.9.1. The flutter_litert release updates Android's CompiledModel runtime to LiteRT Next 2.2.0.
• The example app depends on camera_desktop ^1.2.2.
• No API changes.

4.0.0

• Detections that are not dogs are now dropped instead of returned. In full and poseOnly modes the species classifier's label is checked before a [Dog] is emitted. Previously every animal the body detector found was returned as a Dog, with dog face landmarks run on it, whatever the classifier said. A Dog that is not a dog breaks the guarantee its own type makes, so these are now filtered out. Callers who want every animal regardless of species should use animal_detection directly.

• This is a behaviour change. Code that counted on receiving one result per detected animal will see fewer results. Nothing else about the returned data changed.

• Near-miss classes are recovered rather than lost. The package now ships its own species_mapping.json rather than reading animal_detection's. It maps the domestic block (ImageNet 151-268, 275) plus a wild_canid block (269-274: timber/white/red wolf, coyote, dingo, dhole), whose members are most often a domestic dog the classifier placed on a neighbouring class. Those are returned as species: 'dog'. Every other class resolves to unknown_animal and is dropped, which includes the clothing and object classes a person is most likely to be assigned.

• breed is now null for the near-miss block. The animal is still returned, but the label is withheld rather than naming an animal it probably is not. breed was already null when classification did not run; this adds a third case. See the dartdoc on [Dog.breed].

• Deliberately excluded: hyena (276), which is a feliform rather than a canid. The face landmark model never saw it and, being a regressor with no confidence output, would emit confident but meaningless landmarks with no signal that anything was wrong.

• Added minSpeciesConfidence, an optional second filter on classifier confidence. Defaults to 0.0, meaning off. It is not comparable with cat_detection's value: the classifier is a 1000-class ImageNet model and this is one class's softmax probability, so mass splits across the 125 classes a dog occupies versus a different count for a cat. Tune it against your own imagery.

• faceOnly mode is unchanged. It runs no body detector and no classifier, so no species exists to gate on, and the caller has already asserted the subject.

• Fixed: a cropped cv.Mat passed to detectFromMat returned no detections. Mat.data ignores row stride, so a non-continuous Mat, which is what mat.region(...) returns, was read as though its rows were tightly packed and arrived scrambled. Passing a cropped view produced zero detections or nonsense labels; the same crop with .clone() worked. Non-continuous input is now packed automatically, so no .clone() is needed at the call site. face_detection_tflite, pose_detection and hand_detection already guarded against this; this brings the remaining packages in line.

• Fixed a gap inherited from animal_detection's mapping: ImageNet class 268, Mexican hairless, was absent from the dog block. It resolved to unknown_animal, which was harmless before this release but would now cause that breed to be dropped. It is included here.

• The bundled model files are now explicitly CC BY-NC 4.0, non-commercial use only. The Dart source code remains Apache 2.0 and is unchanged. Only the licensing statement changed: nothing about the weights themselves is different from 3.0.1, and this does not retroactively grant or remove any right. It records the position accurately for the first time. assets/models/dog_face_landmarks_full.tflite and assets/models/dog_face_localizer.tflite are trained on DogFLW, which is CC BY-NC 4.0. The dataset's authors were asked directly how they wanted derived weights licensed, asked for CC BY-NC 4.0 to stay consistent with the source data, and granted permission to publish them on that basis. Using this package in a commercial product runs those weights, which that license does not permit; for commercial use, contact the dataset authors at the Tech4Animals Lab, University of Haifa. See the new NOTICE file.

• Requires animal_detection 4.1.0, which documents its own bundled SuperAnimal body-detection and pose models as academic/non-commercial only and non-transferable. That restriction is independent of the one above: it comes from the Mathis Laboratory's checkpoints rather than from DogFLW. In practice the whole pipeline is non-commercial, by two separate routes, and clearing one would not clear the other.

• The weights are now published on their own at https://huggingface.co/hugocornellier/dog-face-landmarks, alongside a higher-accuracy variant better suited to server-side use, and the training code is public at https://github.com/hugocornellier/dog-face-landmarks-training.

• Depend on flutter_litert ^3.9.0, opencv_dart ^2.2.2, and dartcv4 ^2.3.1. The direct dartcv4 constraint exists only so resolution can never keep a dartcv4 release whose iOS CMake hook hardcodes a 12.0 deployment target, which Xcode 27 rejects; no Dart source imports it.

• Also require hooks ^2.0.0. The dartcv4 2.3.1 link hook uses the hooks 2.x LinkInput API but still accepts hooks 1.x, so a lockfile that kept hooks 1.x failed every profile and release build with a recordedUses compile error. The floor makes pub get move hooks forward (and with it code_assets and objective_c). No Dart source imports it either.

• Raise the floors to Dart 3.10 and Flutter 3.47.5. Earlier Flutter releases pin meta 1.18.0 through flutter_test, which cannot coexist with dartcv4 2.3.1.

• Building for iOS with Xcode 27 needs an iOS 15 deployment target. Set the Runner target (and platform :ios in the Podfile) to 15.0 or newer and add this to the app's pubspec.yaml; hook user-defines are only honoured from the root package, so a dependency cannot supply it for you:

  hooks:
    user_defines:
      dartcv4:
        ios:
          deployment_target: '15.0'
  

  Run flutter clean afterwards so the cached OpenCV build is regenerated.

• Remove the unused direct meta dependency.

• Verified with the hosted animal_detection 4.1.0 and flutter_litert 3.9.0 on macOS 27, Xcode 27, and the iOS 27 simulator.

3.0.1

• Re-exported both face models with static shapes so GPU backends can run them. Trained weights are unchanged; only the export path changed. The previous models were converted with from_keras_model, which leaves the batch dimension dynamic and emits SHAPE / STRIDED_SLICE / PACK in the graph tail. Every GPU backend refuses a graph with dynamic-sized tensors, so both stages ran on CPU on every platform, and switching between Interpreter and CompiledModel changed nothing because both fell back to the same CPU path. Converting from a batch-1 concrete function removes those ops (dog_face_localizer 689 to 597 ops, dog_face_landmarks_full 295 to 283). The landmark model additionally has its deconv ReLU moved out of TRANSPOSE_CONV into a separate RELU op, dropping the opcode from version 4 to 3, which is what lets CompiledModel's GPU accelerator claim the head.

• Output parity against the 3.0.0 models is 0.0 (localizer) and 4.17e-07 (landmarks), so detection quality is unchanged.

• useCompiledModel now defaults to true. The re-exported graphs are accepted by CompiledModel with the default {gpu, cpu} accelerator set, which is the fastest configuration measured on every Apple platform. The existing per-stage try/catch still falls back to the Interpreter if CompiledModel construction fails.

• The landmark stage now defaults to the GPU delegate instead of auto. XNNPACK claims the deconv region with a kernel slower than TFLite's built-in ruy one, so on the re-exported graph XNNPACK is slower than bare CPU. Auto resolves to XNNPACK on Android, macOS, Linux and Windows, so leaving it on auto would have made the landmark stage slower than 3.0.0. Pass landmarkPerformanceConfig to override. Platforms with no GPU delegate fall through to bare CPU, which measures the same as 3.0.0 on this graph.

• Require animal_detection ^3.0.1, which ships its live-camera APIs and updated native example and restores complete pub.dev package analysis.

• Measured on macOS M4 Max, flutter_litert 3.9.0, 25 iterations after 8 warmup, median of sync_p50_ms:

  stage	3.0.0 best	3.0.1 best
  localizer	7.94 ms (XNNPACK)	1.69 ms (CompiledModel {gpu, cpu})
  landmarks	26.64 ms (XNNPACK)	3.85 ms (CompiledModel {gpu, cpu})

3.0.0

• Add opt-in LiteRT Next CompiledModel support to DogDetector.initialize() and the new DogDetector.create(). useCompiledModel defaults to false; accelerators defaults to GPU with CPU fallback and precision to fp32.
• Route the selected backend through the worker isolate and every active stage: animal body detection, species classification, body pose, face localization, and face landmarks. Each compiled graph is numerically verified; unsafe graphs retry on CPU or fall back only that stage to Interpreter.
• Add a macOS full-pipeline parity integration test comparing body boxes and scores, species, every pose point, the face box, and all 46 face landmarks in Interpreter, CompiledModel CPU, and requested GPU+CPU modes.
• Remove the deprecated DogDetectorIsolate. DogDetector has owned its background isolate since 2.0.0 and is now the package's only detector class.
• Require animal_detection ^3.0.0 and keep flutter_litert ^3.8.0.
• Add detectFromCameraFrame() and detectFromCameraImage(), keeping camera pixel conversion, rotation, downscaling, and inference in the detector worker. The native example now matches the face, pose, and hand examples with live camera, still image, and smoothed video-file demos.

2.1.0

• 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.

2.0.0

• Documented DogDetectionMode.faceOnly properly. Behavior is unchanged, but it was described only as "legacy behavior, no SSD". It runs the face localizer on the whole letterboxed image, which is the input the localizer was trained on, and skips the roughly 23MB of body-stage models entirely. The localizer emits a single box, so the mode returns at most one face however many dogs are present, and the returned Dog has no species, breed or pose.

• Removed DogLandmarkModel.ensemble. Swapping the bundled 384px model to MobileNetV3Large left the ensemble mixing backbones: the two downloaded members are still EfficientNetV2S, so the published accuracy figure, measured when all three members shared a backbone, no longer described what the mode actually ran. Re-validating it was not worthwhile. The EfficientNetV2S members measured 156 ms and 242 ms per inference against the bundled model's 83 ms, so the mode cost roughly 960 ms/frame and 109 MB of downloads. Unlike the single bundled model it was never re-measured after the swap, so it was returning results of unknown quality. DogDetector.isEnsembleCached() is removed with it. DogLandmarkModel.full is unchanged and remains the default.

  The v0.0.1-models GitHub release is retained so existing 1.x installations keep working.

• DogDetector now runs the whole pipeline in a background isolate that it owns. initialize() loads the model assets on the main isolate (where rootBundle is available) and transfers them into a worker it spawns, so detection no longer runs on the calling thread. This makes DogDetector the single entry point for the package.

• Deprecated DogDetectorIsolate. It is now a thin delegate to DogDetector and will be removed in the next major release. Migration is a rename: DogDetectorIsolate.spawn(...) becomes DogDetector(...) plus await initialize(), detectDogs becomes detect, and detectDogsFromMat becomes detectFromMat. onDownloadProgress moves from spawn() to initialize().

• DogDetector.detectFromMat now takes imageWidth and imageHeight as optional named arguments, defaulting to the Mat's own cols and rows. Existing call sites that pass them keep working.

• DogDetector.initialize() no longer accepts useIsolateInterpreter, and initializeFromBuffers is no longer part of the public API. The worker isolate owns interpreter creation, so neither had a meaningful effect on the public class. The buffer-based entry point now lives on the internal DogDetectorCore.

• detThreshold is now honored on the isolate path. The previous DogDetectorIsolate never forwarded it to the isolate, so a custom threshold was silently ignored and the pipeline ran at the 0.5 default.

• The example app now uses DogDetector with default (accelerated) performance settings instead of DogDetectorIsolate with PerformanceConfig.disabled.

• Require animal_detection 2.0.0, which replaces its boxed nested input and output tensors with reused flat Float32Lists handed to TFLite as ByteBuffers. Measured on this pipeline over a 3264x2448 photo in profile mode with PerformanceMode.auto, poseOnly drops from 48.2 ms/frame to 15.1 ms, a 3.2x speedup on the shared body pipeline. The full pipeline goes from 452.1 ms/frame to 114.5 ms, though that figure also includes the landmark model swap below rather than the tensor change alone.

• Landmark and pose coordinates shift slightly. animal_detection 2.0.0 fixes ImageUtils.cropAndResize describing an integral crop with pre-truncation floats, which placed landmarks about 0.61px right and 0.52px down of ground truth. Measured over the 311-image CatFLW (measured on cat_detection; the same fix applies here) holdout with real localizer boxes, that cost 0.255 NME_IOD, rising to 1.14 at the 95th percentile, with 72% of images improving. _pipelineVersion is bumped to pipeline_v3 accordingly, so downstream caches re-evaluate stored detections.

• AnimalPoseModel.hrnet now works. animal_detection was requesting superanimal_hrnet_w32_256_float16.tflite while its release publishes superanimal_hrnet_w32_float16.tflite, so selecting HRNet failed with an HTTP 404 on first use and had never worked.

1.5.0

• Replace the bundled dog face landmark model with a MobileNetV3Large backbone (128-channel deconv head) trained on DogFLW at the same 384px input. The asset drops from 54.6 MiB to 11.0 MiB (57.3 MB -> 11.6 MB), a 4.9x reduction, and accuracy improves slightly. Evaluated over all 480 DogFLW test images in absolute image-pixel space at the training crop geometry, NME_IOD is 8.04 versus 8.21 for the previous EfficientNetV2S model. Every facial region is equal or better except mouth (+0.10); both ears, both eyes, nose bridge and nostrils improve. TFLite invoke() also measured about 4x faster on desktop CPU with XNNPACK (134 ms vs 546 ms at 4 threads).

  The input/output signature (float32 [1, 384, 384, 3] -> float32 [1, 92]) is unchanged, so this is a drop-in replacement requiring no caller changes.

  Caveat on the accuracy figures: DogFLW's test split is also used as the validation set during training (early stopping and best-weight restoration monitor it), so both numbers are optimistic in absolute terms. The new model was also given a longer fine-tuning schedule (400 epochs vs 200), so the improvement is not purely architectural.

• Bump the pipeline component of DogDetector.modelVersion to pipeline_v2 so downstream caches holding detections produced by the old model re-evaluate.

• Declare the bundled models' input resolutions as named constants rather than repeating integer literals at each call site. A mismatch between the literal and the bundled model is not reported as an error by the interpreter, which resizes the input tensor and then emits garbage coordinates, so the two call sites for each model could previously drift apart silently.

• DogDetectorIsolate.spawn now defaults performanceConfig to PerformanceMode.auto (Metal on iOS, XNNPACK elsewhere) instead of PerformanceConfig.disabled. The isolate is the documented path for live camera work, so the previous default silently opted the most performance-sensitive callers out of hardware acceleration while the plain DogDetector constructor already defaulted to auto. Measured on the equivalent cat_detection pipeline over a 3264x2448 photo in profile mode, acceleration off ran 1716 ms/frame versus 438 ms/frame with auto, a 3.9x difference. Callers who relied on the old behaviour should pass PerformanceConfig.disabled explicitly.

1.4.0

• Update animal_detection -> 1.4.0, which replaces its shipped 12,944-line SSD anchor table with runtime generation. Detection output is unchanged: verified against the real model over 9 images at 100 runs each with identical detection counts, bit-identical scores, and a worst-case box coordinate delta of 9.3e-05 px. The shared library drops from 15,222 to 2,355 lines and the compiled binary shrinks by about 32 KB.
• Update flutter_litert -> 3.6.0.

1.3.3

• Update flutter_litert -> 3.5.0

1.3.2

• Update flutter_litert -> 3.4.1
• Update animal_detection -> 1.3.2

1.3.1

• Update flutter_litert -> 3.3.1

1.3.0

• Update flutter_litert -> 3.2.0
• Require animal_detection 1.3.0

1.2.3

• Update flutter_litert -> 3.1.1

1.2.2

• Update flutter_litert -> 3.1.0

1.2.1

• Update flutter_litert -> 2.8.3

1.2.0

• Update flutter_litert -> 2.8.0
• Complete Swift Package Manager migration: example apps build via SPM without CocoaPods

1.1.1

• Remove unused Darwin podspecs for Dart-only iOS/macOS plugin registration.
• Require animal_detection 1.1.1.

1.1.0

• Update animal_detection -> 1.1.0
• Update flutter_litert -> 2.5.8

1.0.12

• Update flutter_litert -> 2.5.5

1.0.11

• Update flutter_litert to 2.5.3

1.0.10

• Update flutter_litert -> 2.5.2

1.0.9

• Update flutter_litert -> 2.5.0

1.0.8

• Update flutter_litert -> 2.4.1

1.0.7

• Update flutter_litert -> 2.4.0

1.0.6

• Update flutter_litert -> 2.3.0

1.0.5

• Add public DogDetector.modelVersion and DogDetector.modelVersionFor(...) APIs for downstream cache invalidation.

1.0.4

• Update flutter_litert -> 2.2.0

1.0.3

• Update flutter_litert -> 2.1.0

1.0.2

• Update flutter_litert to 2.0.13
• Update animal_detection to 1.0.2

1.0.1

• Update flutter_litert -> 2.0.12

1.0.0

• First stable release. On-device dog face detection and 46-point facial landmark prediction using TensorFlow Lite. Supports Android, iOS, macOS, Windows, and Linux with automatic hardware acceleration.

0.0.10

• Update documentation

0.0.9

• Update flutter_litert 2.0.8 -> 2.0.10

0.0.8

• Enable auto hardware acceleration by default (XNNPACK on all native platforms, Metal GPU on iOS)
• Update flutter_litert 2.0.6 -> 2.0.8
• Update animal_detection 0.0.5 -> 0.0.6

0.0.7

• Fix Android hang on sequential detect calls

0.0.6

• Fix isolate hanging on sequential detect calls

0.0.5

• Update animal_detection 0.0.3 -> 0.0.4

0.0.4

• Fix Xcode build warnings by declaring PrivacyInfo.xcprivacy as a resource bundle in iOS and macOS podspecs

0.0.3

• Refactor to use shared animal_detection utils

0.0.2

• Added homepage and repository to pubspec.yaml

0.0.1

• Initial release
• Dog face detection with bounding box
• 46 facial landmark extraction (ears, eyes, nose, mouth/chin)
• DogDetector and DogDetectorIsolate APIs
• Support for iOS, Android, macOS, Windows, Linux