PointNet<T>

Models & Types · AiDotNet.PointCloud.Models

Implements the PointNet architecture for processing point cloud data.

Inheritance: NeuralNetworkBase<T> → PointNet<T>

Implements: IPointCloudClassification<T>

For Beginners

PointNet is a pioneering deep learning architecture designed to directly process point clouds.

How It Works

Key innovations of PointNet:

  • Directly processes unordered point sets (no need to convert to voxels or images)
  • Permutation invariant: output doesn’t change if you shuffle the input points
  • Learns both local and global features
  • Uses spatial transformer networks (T-Net) for alignment

Architecture overview:

  1. Input transformation: T-Net learns to align input points
  2. Multi-layer perceptron (MLP): Processes each point independently
  3. Feature transformation: Another T-Net aligns learned features
  4. More MLPs: Further feature extraction
  5. Max pooling: Aggregates information from all points
  6. Global feature vector: Represents the entire point cloud
  7. Classification/Segmentation: Task-specific layers

Why it’s important:

  • First successful deep learning approach for raw point clouds
  • Achieves state-of-the-art results on ModelNet40 classification
  • Foundation for many subsequent point cloud methods
  • Widely used in robotics, autonomous driving, and 3D vision

Reference: “PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation” by Qi et al., CVPR 2017

Example

// Create a PointNet model for 3D object classification with 40 classes
var pointNet = new PointNet<float>(new AiDotNet.Models.Options.PointNetOptions { NumClasses = 40, UseInputTransform = true, UseFeatureTransform = true });
// Process a point cloud (1024 points, 3 coordinates each)
var pointCloud = new Tensor<float>(new[] { 1, 1024, 3 });
Vector<float> classProbabilities = pointNet.ClassifyPointCloud(pointCloud);

Constructors

Constructor Summary
PointNet() Initializes a new instance of the PointNet class with default options.
PointNet(PointNetOptions options, ILossFunction<T> lossFunction, PointNetOptions modelOptions) Initializes a new instance of the PointNet class with configurable options.
PointNet(int numClasses, ILossFunction<T> lossFunction) Initializes a new instance of the PointNet class.

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> ExtractPointFeatures(Tensor<T> pointCloud)
Tensor<T> ForwardWithMemory(Tensor<T> input)
Vector<T> ClassifyPointCloud(Tensor<T> pointCloud)
Vector<T> ExtractGlobalFeatures(Tensor<T> pointCloud)