SetAbstractionLayer<T>

Layers ยท AiDotNet.PointCloud.Models

Implements a Set Abstraction layer for PointNet++.

Inheritance: LayerBase<T> โ†’ SetAbstractionLayer<T>

Implements: IShapeContract

How It Works

For Beginners: Set Abstraction is the core building block of PointNet++.

It performs three operations:

  1. Sampling: Select representative points (centroids)
  2. Grouping: Find neighbors around each centroid
  3. PointNet: Apply mini-PointNet to each local region

Think of it like this:

  • You have 1000 points representing a chair
  • Sampling: Select 100 important points (e.g., corners, edges)
  • Grouping: For each of the 100 points, find nearby points within radius r
  • PointNet: Process each local neighborhood to extract features
  • Result: 100 points with rich features describing local geometry

This creates a hierarchical representation:

  • Input: Many points, basic features (XYZ)
  • Output: Fewer points, rich features (learned patterns)

Constructors

Constructor Summary
SetAbstractionLayer(int numPoints, double searchRadius, int inputChannels, int[] mlpDimensions, int neighborSamples) Single-scale construction, expressed as the one-branch case of the multi-scale form.
SetAbstractionLayer(int numPoints, double[] radii, int inputChannels, int[][] mlpDimensions, int[] neighborSamples) Single-scale construction, expressed as the one-branch case of the multi-scale form.

Properties

Property Summary
Tensor<T> LastCentroidPositions { get; }
bool SupportsTraining { get; }
int OutputChannels { get; }

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
void ClearGradients()
void ResetState()
void UpdateParameters(T learningRate)