DGCNN<T>
Models & Types · AiDotNet.PointCloud.Models
Implements Dynamic Graph CNN (DGCNN) for point cloud processing.
Inheritance: NeuralNetworkBase<T> → DGCNN<T>
Implements: IPointCloudClassification<T>, IPointCloudSegmentation<T>
For Beginners
DGCNN treats point clouds as graphs and uses edge convolutions to learn features.
How It Works
Key innovations of DGCNN:
- Dynamic graph construction: Rebuilds neighborhood graph at each layer based on learned features
- Edge convolution: Learns features from edges connecting nearby points
- Captures local geometric structure more effectively than PointNet
- Adapts to the feature space, not just spatial coordinates
How DGCNN differs from PointNet:
- PointNet: Processes each point independently, then aggregates
- DGCNN: Explicitly models relationships between neighboring points
- Dynamic graphs: Neighborhoods change as features evolve through layers
Edge Convolution explained:
- For each point, find K nearest neighbors (in feature space or spatial)
- Compute edge features: combine point feature with neighbor features
- Apply MLP to edge features
- Aggregate (max pool) edge features for each point
- Result: New features that incorporate local structure
Why dynamic graphs are powerful:
- Early layers: Use spatial proximity (XYZ coordinates)
- Later layers: Use semantic similarity (learned features)
- Example: Points on same chair leg become neighbors even if spatially distant
Architecture:
- Multiple EdgeConv layers with increasing feature dimensions
- Each layer rebuilds k-NN graph based on current features
- Concatenate features from all EdgeConv layers
- Max pooling for global features
- Fully connected layers for classification/segmentation
Applications:
- Classification: Achieves state-of-the-art on ModelNet40
- Part segmentation: Excellent for identifying object parts
- Semantic segmentation: Captures fine-grained geometric details
- Better than PointNet at capturing local structure
Example - chair classification:
- Layer 1: Find spatial neighbors (nearby points)
- Layer 2: Find points with similar low-level features (edges, corners)
- Layer 3: Find points with similar mid-level features (vertical bars, flat surfaces)
- Layer 4: Find points with similar high-level features (legs, back, seat)
- Final: Combine all levels to recognize “chair”
Reference: “Dynamic Graph CNN for Learning on Point Clouds” by Wang et al., ACM Transactions on Graphics 2019
Example
// Create a DGCNN model for point cloud classification with dynamic edge convolution
var dgcnn = new DGCNN<float>(new DGCNNOptions { NumClasses = 40, EdgeConvChannels = new[] { 64, 64, 128, 256 }, KnnK = 20, UseDropout = true, DropoutRate = 0.5 });
Constructors
| Constructor | Summary |
|---|---|
DGCNN() |
Initializes a new instance of the DGCNN class with default options. |
DGCNN(DGCNNOptions options, ILossFunction<T> lossFunction) |
Initializes a new instance of the DGCNN class with configurable options. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> ExtractPointFeatures(Tensor<T> pointCloud) |
|
Tensor<T> ForwardForTraining(Tensor<T> input) |
Routes the tape-based training forward through ForwardWithMemory — the DGCNN architecture (multi-scale EdgeConv skip-concatenation → global max-pool → head) is NOT a plain sequential layer stack, so the base’s default sequential `ForwardF… |
Tensor<T> ForwardWithMemory(Tensor<T> input) |
|
Tensor<T> SegmentPointCloud(Tensor<T> pointCloud) |
|
Vector<T> ClassifyPointCloud(Tensor<T> pointCloud) |
|
Vector<T> ExtractGlobalFeatures(Tensor<T> pointCloud) |