KMaXDeepLab<T>

Models & Types · AiDotNet.ComputerVision.Segmentation.Panoptic

kMaX-DeepLab: k-means Mask Transformer for panoptic segmentation.

Inheritance: NeuralNetworkBase<T> → SegmentationModelBase<T> → PanopticSegmentationBase<T> → KMaXDeepLab<T>

For Beginners

Panoptic segmentation (stuff + things). Scene understanding for autonomous driving.

Common use cases:

  • Panoptic segmentation (stuff + things)
  • Scene understanding for autonomous driving
  • Robotics environment parsing
  • Dense scene labeling

How It Works

Technical Details:

  • k-means cross-attention replacing standard cross-attention: every pixel is assigned to its argmax cluster and

the assigned pixel values update the cluster centers

  • Each decoder stage: k-means cross-attention, self-attention among the centers and an FFN, each with learnable

projections, residual connections and BatchNorm (deeplab2’s kMaXTransformerLayer)

  • Trained with the PQ-style loss, the mask-ID cross-entropy, the pixel-wise instance discrimination loss and an

auxiliary semantic loss, every decoder stage supervised (weights in KMaXDeepLabOptions)

Reference: Yu et al., “k-means Mask Transformer”, ECCV 2022.

Example

// Create a kMaX-DeepLab model for panoptic segmentation
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.ThreeDimensional,
    taskType: NeuralNetworkTaskType.MultiClassClassification,
    inputHeight: 512, inputWidth: 512, inputDepth: 3, outputSize: 133);
var model = new KMaXDeepLab<double>(architecture);
// Every tunable is now set through the options object; these are the defaults:
var custom = new KMaXDeepLab<double>(architecture,
    options: new KMaXDeepLabOptions { NumClasses = 133, DropRate = 0.1, ModelSize = KMaXDeepLabModelSize.R50 });
            
// Or load a pre-trained ONNX model for autonomous driving scene parsing
var onnxModel = new KMaXDeepLab<double>(architecture, "kmaxdeeplab.onnx");

Constructors

Constructor Summary
KMaXDeepLab(NeuralNetworkArchitecture<T> architecture, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction, KMaXDeepLabOptions options) Initializes KMaXDeepLab in native (trainable) mode.
KMaXDeepLab(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, KMaXDeepLabOptions options) Initializes KMaXDeepLab in ONNX (inference-only) mode.

Methods

Method Summary
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
ModelMetadata<T> GetModelMetadata() Collects metadata describing this model’s configuration.
PanopticSegmentationResult<T> SegmentPanoptic(Tensor<T> image)
Tensor<T> ForwardForTraining(Tensor<T> input) The training forward for a caller-supplied loss: the same semantic scores Forward returns (the layers are not a sequential chain, so the base layer walk would train a different tensor).