EdgeConvLayer<T>
Layers · AiDotNet.PointCloud.Models
Implements an Edge Convolution layer for DGCNN.
Inheritance: LayerBase<T> → EdgeConvLayer<T>
Implements: IShapeContract
How It Works
For Beginners: EdgeConv is the key building block of DGCNN.
What EdgeConv does:
- Build a graph: Connect each point to its K nearest neighbors
- Compute edge features: For each edge, combine features from both endpoints
- Transform: Apply MLP to edge features
- Aggregate: Max pool edge features for each point
Edge feature computation:
- For point i with neighbor j: edge_ij = MLP(concat(feature_i, feature_j - feature_i))
- This captures both the point’s feature and its relationship to neighbors
Example:
- Point at chair leg: High features, neighbors also high (smooth region)
- Point at edge: High features, neighbors vary (captures edge information)
- The difference (feature_j - feature_i) captures local geometry
Dynamic aspect:
- Early layers: Neighbors are spatially close points
- Later layers: Neighbors are semantically similar points
- Graph structure adapts as features evolve
Constructors
| Constructor | Summary |
|---|---|
EdgeConvLayer(int inputChannels, int outputChannels, int k) |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<ILayer<T>> GetSubLayers() |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
|
void ClearGradients() |
|
void ResetState() |
|
void UpdateParameters(T learningRate) |