SelfieSegmentation class

Performs selfie/person segmentation using MediaPipe TFLite models.

Generates a per-pixel probability mask separating foreground (person) from background. Works on full images - no face detection required.

See the official model cards for architecture details, training data, and intended use cases:

Model Variants

Three model variants are available via SegmentationConfig.model:

Platform Support

All model variants work on all platforms. Binary models (general/landscape) use the Convolution2DTransposeBias custom op which is bundled for each platform.

Platform Delegate Notes
iOS Metal GPU Custom ops statically linked via CocoaPods
Android GPU/CPU Custom ops built via CMake, loaded at runtime
macOS CPU Custom ops bundled as dylib
Linux CPU Custom ops bundled as .so
Windows CPU Custom ops bundled as dll

Example

// Default: fast binary segmentation (~244KB model)
final segmenter = await SelfieSegmentation.create();
final mask = await segmenter(imageBytes);
final binary = mask.toBinary(threshold: 0.5);
segmenter.dispose();

// Multiclass: per-class body part segmentation (~16MB model)
final multiSeg = await SelfieSegmentation.create(
  config: SegmentationConfig(model: SegmentationModel.multiclass),
);
final multiMask = await multiSeg(imageBytes);
if (multiMask is MulticlassSegmentationMask) {
  final hair = multiMask.hairMask;
}
multiSeg.dispose();

Memory Considerations (varies by model)

  • General/Landscape: ~244KB model + ~768KB buffers
  • Multiclass: ~16MB model + ~2.3MB buffers

Integration with FaceDetector

For combined face detection and segmentation, use FaceDetector:

final detector = FaceDetector();
await detector.initialize();
await detector.initializeSegmentation();

final faces = await detector.detectFacesFromBytes(imageBytes);
final mask = await detector.getSegmentationMask(imageBytes);

Properties

config → SegmentationConfig
Current configuration.
no setter
hasGpuDelegateFailed → bool
Whether GPU delegate failed during inference.
no setter
hashCode → int
The hash code for this object.
no setterinherited
inputHeight → int
Model input height in pixels.
no setter
inputWidth → int
Model input width in pixels.
no setter
isDisposed → bool
Whether this instance has been disposed.
no setter
model → SegmentationModel
The segmentation model variant in use.
no setter
outputChannels → int
Number of output channels (1 for binary models, 6 for multiclass).
no setter
outputHeight → int
Output mask height.
no setter
outputWidth → int
Output mask width (may differ from input due to model architecture).
no setter
runtimeType → Type
A representation of the runtime type of the object.
no setterinherited

Methods

call(Mat image, {Float32List? buffer}) → Future<SegmentationMask>
Segments an image to separate foreground (person) from background.
callFromBytes(Uint8List imageBytes, {Float32List? buffer}) → Future<SegmentationMask>
Segments an image from encoded bytes (JPEG, PNG, etc.).
dispose() → void
Releases all TensorFlow Lite resources held by this model.
disposeAsync() → Future<void>
Asynchronously releases all resources, ensuring the background isolate is fully terminated before freeing the native interpreter.
noSuchMethod(Invocation invocation) → dynamic
Invoked when a nonexistent method or property is accessed.
inherited
toString() → String
A string representation of this object.
inherited

Operators

operator ==(Object other) → bool
The equality operator.
inherited

Static Methods

create({SegmentationConfig config = const SegmentationConfig()}) → Future<SelfieSegmentation>
Creates and initializes a selfie segmentation model instance.
createCompiledFromBuffer(Uint8List modelBytes, {SegmentationConfig config = const SegmentationConfig(), Set<Accelerator> accelerators = const {Accelerator.gpu, Accelerator.cpu}, Precision precision = Precision.fp16}) → Future<SelfieSegmentation>
Creates a selfie segmentation model backed by LiteRT CompiledModel.
createFromBuffer(Uint8List modelBytes, {SegmentationConfig config = const SegmentationConfig()}) → Future<SelfieSegmentation>
Creates a selfie segmentation model from pre-loaded model bytes.