TNetLayer<T>

Layers · AiDotNet.PointCloud.Layers

Implements a Transformation Network (T-Net) for learning spatial transformations of point clouds.

Inheritance: LayerBase<T> → TNetLayer<T>

Implements: IShapeContract

How It Works

For Beginners: T-Net learns to align and normalize point clouds before processing.

What T-Net does:

  • Learns a transformation matrix to apply to input points
  • Aligns point clouds to a canonical orientation
  • Makes the network more robust to rotations and translations
  • Helps the network focus on shape rather than orientation

How it works:

  1. Takes point cloud as input
  2. Processes it through small neural network
  3. Outputs a transformation matrix (e.g., 3x3 for spatial, KxK for feature)
  4. Applies this matrix to transform the input

Two types of T-Net in PointNet:

  • Input T-Net: 3x3 matrix to align XYZ coordinates
  • Feature T-Net: KxK matrix to align high-dimensional features

Benefits:

  • Achieves invariance to rigid transformations
  • Normalizes point cloud orientation
  • Improves classification and segmentation accuracy

Example:

  • Input: Point cloud that might be rotated randomly
  • T-Net learns: “Rotate this cloud 45 degrees to align it”
  • Output: Aligned point cloud in standard orientation

Constructors

Constructor Summary
TNetLayer(int transformDim, int numFeatures, int[] mlpChannels, int[] fcChannels) Initializes a new instance of the TNetLayer class.

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) Derived from this type’s [TensorLayout] declarations.
void ClearGradients()
void ResetState()
void UpdateParameters(T learningRate)
void UpdateParameters(Vector<T> parameters)