PointTransformerV3<T>

Models & Types · AiDotNet.ComputerVision.Segmentation.PointCloud

Point Transformer V3: Simpler, Faster, Stronger – per-point semantic segmentation of 3D point clouds.

Inheritance: NeuralNetworkBase<T> → SegmentationModelBase<T> → SemanticSegmentationBase<T> → PointTransformerV3<T>

For Beginners

Give the model a point cloud – a list of points, each with its xyz position and a few extra values such as colour and surface normal – and it returns a score for every class at every point. It is used for LiDAR and indoor-scan understanding.

How It Works

Architecture (Wu et al., CVPR 2024; Pointcept PT-v3m1), built by PointTransformerV3Encoder:

Input: [points, 3 + InputChannels] or [batch, points, 3 + InputChannels], xyz in metres first. Output: class logits [..., points, NumClasses].

Example

var architecture = new NeuralNetworkArchitecture<float>(
    InputType.TwoDimensional, NeuralNetworkTaskType.ImageSegmentation,
    inputSize: 9, outputSize: 20);
var options = new PointTransformerV3Options { NumClasses = 20 };
var model = new PointTransformerV3<float>(architecture, options: options);
            
var cloud = Tensor<float>.CreateRandom(4096, 9);   // xyz, colour, normal
Tensor<float> logits = model.Predict(cloud);          // [4096, 20]

Constructors

Constructor Summary
PointTransformerV3(NeuralNetworkArchitecture<T> architecture, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction, PointTransformerV3Options options) Initializes PointTransformerV3 in native (trainable) mode.
PointTransformerV3(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, PointTransformerV3Options options) Initializes PointTransformerV3 in ONNX (inference-only) mode.

Methods

Method Summary
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
ModelMetadata<T> GetModelMetadata()
Tensor<T> GetClassMap(Tensor<T> image) Argmax over the class (last) axis: one class index per point.
Tensor<T> GetProbabilityMap(Tensor<T> image) Softmax over the class (last) axis.