pose_detection changelog | Flutter package

3.6.0 #

iOS now runs the YOLO person detector on the Metal GPU on the default CompiledModel path, cutting live-camera latency roughly 4x on device (measured on iPhone: ~100ms to ~22ms per frame, with detection keeping pace with the 24 fps camera stream). The CPU pin dated from v2.0.6, when the old TFLite Metal GPU delegate produced inconsistent detection counts (10 vs 2 on the same image); the LiteRT Next Metal accelerator used by the CompiledModel path shows stable counts at fp16 on device. The interpreter path keeps its XNNPACK override on iOS because it still uses the old delegate.
CompiledModel GPU-to-CPU fallbacks are no longer silent: the YOLO and landmark models log the fallback error and the compiled accelerator set in debug builds ([pose-accel] lines).
Fix (example): PoseSmoother rebuilt each Pose to write filtered landmark coordinates but omitted segmentationMask from the constructor call, so enabling smoothing silently dropped masks. The example app never enables segmentation so nothing surfaced it, but the smoother reads as general-purpose and example code exists to be copied. Adds example/test with a regression test that fails without the fix, plus coverage for the null-mask case, the unfiltered scalar fields, and the disabled no-op path.
Deprecate CameraPoseOverlayPainter in favour of CameraPoseOverlay. The painter scales x and y independently, so it only maps correctly when the paint box already matches the detection image's aspect ratio; over a cover-fitted preview it stretches the overlay. The widget performs the same job through a cover fit. Deprecated rather than changed, since altering its mapping would move overlays for downstream users who sized their box to match.
Adopt the shared flutter_litert 3.6.0 helpers in place of local copies: aggregateActiveAccelerator for the multi-runner accelerator aggregation (this package already had the correct any-WebGPU behaviour, so this is de-duplication rather than a fix), compiledModelFromBufferAuto for the {gpu, cpu} branch at both CompiledModel call sites, and iouLTRB for track matching.
Remove the web_image_utils re-export shim and import from flutter_litert directly.
Update flutter_litert -> 3.6.0.
Expand the README live camera section with the full production pipeline (frame throttling, orientation handling, cover-fit overlay mapping).

3.5.3 #

Fix landmark confidence decoding for the lite, full, and heavy BlazePose models. Their overall score tensor is already produced by a TFLite LOGISTIC op, so it is now used directly instead of passing through a second sigmoid. The default minLandmarkScore: 0.5 gate can once again reject blank or low-confidence crops.
Return the bundled YOLOv8 person's actual class probability as Pose.score. The detector now uses flutter_litert's explicit probability mode rather than compensating for its legacy logit decoder by transforming the threshold. Filtering is unchanged, while reported confidence is calibrated correctly.
Bump the pose pipeline cache key because identical inputs can now produce different landmark inclusion and confidence output.
Update flutter_litert -> 3.5.1.

3.5.2 #

Update flutter_litert -> 3.5.0

3.5.1 #

Web: the auto accelerator now resolves through flutter_litert's capability probe (resolveWebAccelerator): WebGPU is selected only on Chromium with a real hardware adapter, and everything else starts on WASM. Fixes Firefox 152, whose WebGPU compiles and runs cleanly but far slower than WASM SIMD, so the error-driven fallback could never catch it.
Web: after an auto init lands on WebGPU, a timed warmup on the YOLO stage (WebGpuFallback.maybeSwapIfWebGpuSlow) swaps all runners to WASM when the median run exceeds the budget.
Web: Safari initializes again; flutter_litert now serves LiteRT.js from its wasm directory default, so Safari receives the compat build instead of failing to parse the relaxed-SIMD build (relaxed simd instructions not supported).
Update flutter_litert -> 3.4.1 (also brings the web CompiledModel WebGPU compile watchdog: a compile attempt that never settles falls back to WASM instead of hanging).
example_web: mirror backend + fps into the page title and add ?screen=camera|video|still deep links for automated validation.

3.5.0 #

Add opt-in person segmentation masks. Initialize with enableSegmentation: true (under PoseMode.boxesAndLandmarks) to receive a coarse BlazePose person mask per detected pose via Pose.segmentationMask. The landmark model already computes this mask on every inference, so returning it adds no inference cost. New SegmentationMask type exposes confidenceAt(x, y) sampling and toRgbaBytes() for rendering. Off by default; populated on native platforms only (web accepts the flag for API parity but leaves the mask null).
Fix the YOLOv8 person detector emitting phantom detections. The bundled model outputs class probabilities, not raw logits, but the shared decoder applies a sigmoid to every score, so background anchors near 0 cleared the confidence gate as sigmoid(0) = 0.5. That produced thousands of low-confidence boxes that NMS then collapsed into a random one-to-four detections. The detector now cancels the extra activation so the threshold compares against the model's actual probability.

3.4.0 #

Update flutter_litert -> 3.2.0
Import native-only flutter_litert APIs via package:flutter_litert/native.dart so they resolve under static analysis (flutter_litert 3.2.0 moved InterpreterFactory, InterpreterPool, and IsolateWorkerBase behind the native conditional export). No runtime or API change.

3.3.0 #

Update flutter_litert -> 3.1.1
Add optional LiteRT Next CompiledModel inference (GPU with automatic CPU fallback); enable with useCompiledModel: true. The default engine remains the Interpreter, so existing code is unchanged.
Add a flat YOLO decoder and adopt the shared flutter_litert helpers (flat YOLO decode, GPU fallback, camera frame decode plan).

3.2.1 #

Update flutter_litert -> 2.8.3

3.2.0 #

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

3.1.5 #

Remove unused Darwin podspecs for Dart-only iOS/macOS plugin registration.

3.1.4 #

Update flutter_litert -> 2.5.8

3.1.3 #

Update flutter_litert -> 2.5.5

3.1.2 #

Update flutter_litert to 2.5.3 and camera_desktop to 1.1.4

3.1.1 #

Update documentation and dartdocs

3.1.0 #

Web: LiteRT.js (WebGPU + WASM fallback) is now the default web runtime.

useLiteRt now defaults to true so no opt-in is required. Pass useLiteRt: false to use the legacy tflite-js path. liteRtAccelerator (String, default 'auto') controls the backend: 'auto' and 'webgpu' request WebGPU with WASM fallback when WebGPU compile fails, while 'wasm' opts out of GPU. On browsers without WebGPU support, 'auto' falls back to WASM transparently. Both runtimes load from CDN on first use with no web/index.html changes required; see configureLiteRtLoader in flutter_litert for self-hosting options.
Update flutter_litert -> 2.5.2

3.0.2 #

Update flutter_litert -> 2.5.0

3.0.1 #

Update flutter_litert -> 2.4.1

3.0.0 #

Breaking:

PoseDetector configuration moves from the constructor to initialize(). PoseDetector({mode: ..., landmarkModel: ..., ...}) → PoseDetector() + await detector.initialize(mode: ..., landmarkModel: ..., ...). Matches FaceDetector's shape. PoseDetector.create({...}) continues to accept the same named params unchanged.
PoseDetector.detectFromMat signature changed. detectFromMat(cv.Mat, {required int imageWidth, required int imageHeight}) is now detectFromMat(cv.Mat). Dimensions are read from the Mat directly. Existing callers must drop the imageWidth and imageHeight named arguments.
detect(...) no longer swallows exceptions. Undecodable native image bytes now throw FormatException (matching FaceDetector and HandDetector behaviour) rather than silently returning an empty list. Wrap detect(...) in a try/catch if your callsite depended on the previous silent-failure behaviour. On web, decode failure still returns an empty list because browser image decode failure does not throw through this API.
On native platforms, inference now runs in a dedicated background isolate, keeping the UI thread free. Previously, native inference ran on the calling thread.
Add PoseDetector.create({...}) one-step factory (mirrors FaceDetector.create and HandDetector.create).
Add detectFromFilepath(String path): reads the file and delegates to detect.
Add detectFromMatBytes(Uint8List, {required int width, required int height, int matType}) zero-copy fast path via TransferableTypedData.
Add initializeFromBuffers({required Uint8List yoloBytes, required Uint8List landmarkBytes}) for callers that load model bytes independently of Flutter's asset system.
Add isReady getter as an alias for isInitialized.
Add a top-level Live Camera Detection section to the README, modelled on face_detection_tflite's packYuv420 + native cv.cvtColor pattern, and remove the orphan assets/models/pose_detection.tflite left over from the pre-YOLOv8n scaffold.
Expand flutter_litert re-exports through the pose_detection barrel to match face_detection_tflite: tensor helpers (createNHWCTensor4D, fillNHWC4D, allocTensorShape, flattenDynamicTensor), math helpers (sigmoid, sigmoidClipped, clamp01, clip), letterbox helpers (computeLetterboxParams, LetterboxParams), BGR→RGB byte helpers (bgrBytesToRgbFloat32, bgrBytesToSignedFloat32), and PerformanceMode. Consumers no longer need a direct flutter_litert import for these.

2.1.1 #

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

2.1.0 #

Fix live camera in the example app on Android (previously detections were sideways and unreliable):
- Apply rotation to raw landscape camera frames before detection so the pose detector sees upright people. _rotationFlagForFrame handles all four device orientations (portrait up/down, landscape left/right) via a combined sensorOrientation + DeviceOrientation formula.
- Mirror the overlay on Android front camera to match CameraPreview's auto-mirrored preview texture.
- Replace the per-pixel Dart YUV loop with flutter_litert's shared packYuv420 helper + native cv.cvtColor on mobile (iOS NV12, Android NV21 / I420).
- Replace the per-pixel Dart BGRA→BGR / RGBA→BGR loop with native cv.cvtColor on desktop (macOS / Linux).
Align example app live-camera layout with face_detection_tflite: Material+Row top bar (replaces AppBar), flip-camera button, FPS + detection-time display, rotating top bar in landscape with safe-area padding, and a settings popup housing pose-specific controls (landmark-model chips: Lite / Full / Heavy).
Re-export packYuv420, YuvPlane, YuvLayout, and PackedYuv from flutter_litert through the pose_detection barrel.
Update flutter_litert to ^2.2.0.

2.0.10 #

Update flutter_litert -> 2.1.0

2.0.9 #

Update flutter_litert to 2.0.13

2.0.8 #

Update flutter_litert -> 2.0.12

2.0.7 #

Update flutter_litert 2.0.10 -> 2.0.11

2.0.6 #

Fixed Metal GPU delegate producing inconsistent detection counts on iOS

2.0.5 #

Update documentation

2.0.4 #

Update flutter_litert 2.0.8 -> 2.0.10

2.0.3 #

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

2.0.2 #

Update flutter_litert 2.0.5 -> 2.0.6

2.0.1 #

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

2.0.0 #

Breaking: Point now uses double coordinates. BoundingBox is now a 4-corner Point-based type.

Use shared Point and BoundingBox from flutter_litert 2.0.0
toPixel() now returns full-precision double coordinates (was truncating to int)
Extract PersonDetectorBase shared between native and web detectors
Simplify model classes and detector implementations
Remove integration tests from unit test suite
Remove dead test helpers (test_config.dart)

1.0.7 #

Update camera_desktop 1.0.1 -> 1.0.3

1.0.6 #

Update flutter_litert -> 1.2.0
Refactor to use flutter_litert shared utilities (InterpreterFactory, InterpreterPool, PerformanceConfig)

1.0.5 #

Update opencv_dart 2.1.0 -> 2.2.1
Update flutter_litert 1.0.2 -> 1.0.3

1.0.4 #

Update flutter_litert 1.0.1 -> 1.0.2

1.0.3 #

Update documentation

1.0.2 #

Update flutter_litert to 1.0.1, camera to 0.12.0

1.0.1 #

Update flutter_litert to 0.2.2

1.0.0 #

Initial release
Person detection using YOLOv8
Pose landmark detection with MediaPipe Pose
Support for lite, full, and heavy models
Box-only and full landmark detection modes

pose_detection 3.6.0 pose_detection: ^3.6.0 copied to clipboard

Metadata