MaskRCNN<T>

Models & Types · AiDotNet.ComputerVision.Segmentation.InstanceSegmentation

Mask R-CNN for instance segmentation.

Inheritance: ModelBase<T, Tensor<T>, Tensor<T>> → InstanceSegmenterBase<T> → MaskRCNN<T>

Implements: IGeneratedParameterRegistrar<T>

For Beginners

Mask R-CNN extends Faster R-CNN by adding a mask prediction branch parallel to the box classification and regression branches. It’s a two-stage detector that first proposes regions, then classifies them and predicts masks.

How It Works

Key features:

  • Two-stage detection with RPN and RoI heads
  • Parallel mask prediction branch
  • RoIAlign for precise spatial alignment
  • Decoupled mask and class prediction

Reference: He et al., “Mask R-CNN”, ICCV 2017

Constructors

Constructor Summary
MaskRCNN(InstanceSegmentationOptions<T> options) Creates a new Mask R-CNN model.

Properties

Property Summary
string Name { get; }

Methods

Method Summary
InstanceSegmentationResult<T> Segment(Tensor<T> image)
Task LoadWeightsAsync(string pathOrUrl, CancellationToken cancellationToken)
long GetParameterCount()
void SaveWeights(string path)
void TrainInstances(Tensor<T> input, IReadOnlyList<InstanceSegmentationTrainingTarget<T>> targets) Trains every branch with the Mask R-CNN multi-task loss L = L_cls + L_box + L_mask, plus the RPN’s proposal loss (He et al.