mp_vision

Vision task APIs, running modes, result types, and mask ownership.

Task API#

Task	Image method	Video method	Live method	Result
`FaceDetector`	`detect`	`detectForVideo`	`detectAsync`	`DetectionResult`
`FaceLandmarker`	`detect`	`detectForVideo`	`detectAsync`	`FaceLandmarkerResult`
`GestureRecognizer`	`recognize`	`recognizeForVideo`	`recognizeAsync`	`GestureRecognizerResult`
`HandLandmarker`	`detect`	`detectForVideo`	`detectAsync`	`HandLandmarkerResult`
`HolisticLandmarker`	`detect`	`detectForVideo`	`detectAsync`	`HolisticLandmarkerResult`
`ImageClassifier`	`classify`	`classifyForVideo`	`classifyAsync`	`ClassificationResult`
`ImageEmbedder`	`embed`	`embedForVideo`	`embedAsync`	`EmbeddingResult`
`ImageSegmenter`	`segment`	`segmentForVideo`	`segmentAsync`	`ImageSegmenterResult`
`InteractiveSegmenter`	`segment`	—	—	`ImageSegmenterResult`
`ObjectDetector`	`detect`	`detectForVideo`	`detectAsync`	`DetectionResult`
`PoseLandmarker`	`detect`	`detectForVideo`	`detectAsync`	`PoseLandmarkerResult`

Every task is created with TaskName.create(TaskNameOptions(...)) and closed with close(). Classification, embedding, landmark, detection, mask, and matrix containers come from mp_core and are immutable. Live results are read from the task's results stream as VisionLiveResult<T> values.

Option groups#

All option classes require BaseOptions and, except for InteractiveSegmenterOptions, accept a VisionRunningMode. Detector and landmarker options expose task-specific confidence thresholds and result counts. Classifier-based tasks use ClassifierOptions; embedders use EmbedderOptions.

Mask and optional-output switches are disabled or conservative by default:

FaceLandmarkerOptions can return blendshapes and facial transformation matrices.
HolisticLandmarkerOptions can return face blendshapes and pose segmentation masks.
PoseLandmarkerOptions can return pose segmentation masks.
ImageSegmenterOptions returns confidence masks by default and can return a category mask.

ImageProcessingOptions supplies a clockwise rotation in multiples of 90 degrees and an optional normalized region of interest for one call. Task construction options remain unchanged between calls.

Running modes#

Construct the task with exactly one VisionRunningMode. Calling an image method on a video task, reusing a timestamp, or submitting a frame after close fails immediately in Dart.

The web and current native adapters serialize live-stream submissions over the upstream video APIs, then emit copied results in timestamp order. Native C callbacks remain disabled until the bridge can copy callback-owned memory safely from native worker threads.

Segmentation masks#

Native and web mask storage is copied into MpImage before the upstream result is closed. Keep mask output disabled when it is not needed; full-resolution floating-point masks can dominate result size.