PointNetPlusPlus<T>

Models & Types · AiDotNet.PointCloud.Models

Implements the PointNet++ architecture for hierarchical point cloud processing.

Inheritance: NeuralNetworkBase<T> → PointNetPlusPlus<T>

Implements: IPointCloudClassification<T>, IPointCloudSegmentation<T>

For Beginners

PointNet++ extends PointNet by adding hierarchical feature learning at multiple scales.

How It Works

Key improvements over PointNet:

  • Hierarchical structure: Processes point clouds at multiple resolutions
  • Local context: Captures fine-grained local patterns
  • Multi-scale grouping: Learns features at different scales simultaneously
  • Better generalization: More robust to non-uniform point density

Architecture components:

  1. Set Abstraction Layers: Hierarchically group points and extract features
  • Sampling: Select subset of points as centroids
  • Grouping: Find neighboring points around each centroid
  • PointNet layer: Extract features from each local region
  1. Feature Propagation Layers: Upsample features for segmentation tasks
  • Interpolation: Propagate features from coarse to fine levels
  • Skip connections: Combine with features from encoder

Why hierarchical learning matters:

  • Different patterns exist at different scales (like edges vs. shapes in images)
  • Local context provides detailed geometry information
  • Global context provides overall shape understanding
  • Combining both gives comprehensive understanding

Applications:

  • Fine-grained classification
  • Part segmentation (identifying specific parts of objects)
  • Semantic segmentation (labeling each point)
  • Better performance on complex, detailed shapes

Example use case - autonomous driving:

  • Coarse level: Identify general object shapes (car, pedestrian)
  • Medium level: Recognize object parts (wheels, windows)
  • Fine level: Detect details (door handles, mirrors)

Reference: “PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space” by Qi et al., NeurIPS 2017

Example

// Create a PointNet++ model for hierarchical point cloud classification
var options = new AiDotNet.Models.Options.PointNetPlusPlusOptions
{
    NumClasses = 40,
    UseMultiScaleGrouping = true
};
var pointNetPP = new PointNetPlusPlus<float>(options);

Constructors

Constructor Summary
PointNetPlusPlus() Initializes a new instance of the PointNetPlusPlus class with default options.
PointNetPlusPlus(PointNetPlusPlusOptions options, ILossFunction<T> lossFunction, PointNetPlusPlusOptions modelOptions, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Initializes a new instance of the PointNetPlusPlus class with configurable options.
PointNetPlusPlus(int numClasses, int[] samplingRates, double[] searchRadii, int[][] mlpDimensions, ILossFunction<T> lossFunction) Initializes a new instance of the PointNetPlusPlus 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)
Tensor<T> SegmentPointCloud(Tensor<T> pointCloud)
Vector<T> ClassifyPointCloud(Tensor<T> pointCloud)
Vector<T> ExtractGlobalFeatures(Tensor<T> pointCloud)