Source-linked AI summary
A Graph-CNN for 3D Point Cloud Classification
Yingxue Zhang, Michael Rabbat
TL;DR
3D point-cloud classification must handle unordered points and irregular graph structure while retaining local geometry. PointGCN addresses this with localized graph convolutions and global or multi-resolution pooling, achieving competitive ModelNet performance and lower variation than competing point-based methods. The paper presents graph-based local structure as a practical route to stable classification without requiring multiple input views.
Problem
3D point-cloud classification requires handling point sets while preserving local geometric context that point-wise or voxel-based approaches may inadequately represent.
Method
PointGCN combines localized graph convolutions with two point-cloud-specific pooling operations to summarize graph signals across receptive fields.
Results
PointGCN achieves average classification accuracy comparable to the state of the art on ModelNet, with substantially lower variance than existing point-based methods.
Takeaways & Limitations
Graph structure supplies local geometric information, while the proposed architecture remains robust to point ordering and rotation in the reported ModelNet setting.
Takeaways & Limitations
The approach assumes that K-localized graph convolution can explore features from different receptive fields.
Abstract
from arXiv · showhide
Graph convolutional neural networks (Graph-CNNs) extend traditional CNNs to handle data that is supported on a graph. Major challenges when working with data on graphs are that the support set (the vertices of the graph) do not typically have a natural ordering, and in general, the topology of the graph is not regular (i.e., vertices do not all have the same number of neighbors). Thus, Graph-CNNs have huge potential to deal with 3D point cloud data which has been obtained from sampling a manifold. In this paper, we develop a Graph-CNN for classifying 3D point cloud data, called PointGCN. The architecture combines localized graph convolutions with two types of graph downsampling operations (also known as pooling). By the effective exploration of the point cloud local structure using the Graph-CNN, the proposed architecture achieves competitive performance on the 3D object classification benchmark ModelNet, and our architecture is more stable than competing schemes.
1. INTRODUCTION
Graph-CNNs adapt convolutional learning to graph-supported data, addressing irregular topology and unordered vertices. PointGCN applies localized graph convolutions and specialized pooling to 3D point-cloud classification, achieving competitive ModelNet accuracy with lower variance.
- Background: Graph-CNNs extend CNN-style processing to data represented on graph vertices or edges.Graph signal processing provides filtering and related operations for such data.
- Related work: Existing graph convolutions use spectral filters, localized Chebyshev-polynomial operations, or spatial random-walk receptive fields.Graph multiresolution clustering and pooling are also established components of prior approaches.
- Related work: Graph kernels address graph classification rather than signal processing on graphs and can have quadratic training complexity in the number of graphs.
- 3D point clouds: Graph-based processing represents each point-cloud sample with one vertex per point and edges between nearby points.This approach has supported visualization, in-painting, and compression of 3D point clouds.
- PointGCN: PointGCN classifies 3D point clouds by exploring local graph structure with graph convolutions and two point-cloud-specific pooling types.It learns latent signatures at different receptive fields while both graph structure and signals vary across inputs.
- Results: PointGCN reaches average ModelNet accuracy comparable to the state of the art while showing substantially lower variance than existing point-based methods.
2. PROBLEM STATEMENT
The paper formulates point-cloud classification as predicting a finite class label for an unseen set of 3D points. It motivates graph-based processing as a way to preserve local geometric context that point-wise methods may lose.
- Problem formulation: The task uses m labeled training instances, each pairing an input point set with an output class label.Each input contains n points in R3, and the goal is to predict the class of a new unseen input.
- Existing approaches: Prior methods classify point clouds through multi-view image rendering or voxelization, with preprocessing, computation, discretization, or sparsity drawbacks.
- Existing approaches: PointNet and Deep Sets learn permutation-invariant global representations, but limited local-structure exploration can lose each point’s local context.PointNet++ forms local clusters, yet still treats points individually within those clusters.
- Graph-based formulation: The proposed graph-based formulation encodes local geometric details through pairwise distances between a center point and its neighbors.
3. METHODOLOGY
PointGCN represents each point cloud as a weighted nearest-neighbor graph, applies localized graph filters, and combines global or multi-resolution pooling with fully connected classification. Laplacian rescaling, localized receptive fields, and pooling operations form the main architectural components.
- Graph construction: Each point cloud is represented by a symmetrized k-nearest-neighbor graph with Gaussian-weighted edges, while its three coordinate dimensions serve as graph signals.
- Graph convolution: Graph filters transform signals using the normalized Laplacian and can be expressed with Chebyshev polynomials whose learned coefficients define graph convolutional filters.
- Graph convolution: Rescaling the Laplacian places all eigenvalues in [−1, 1], improving learning stability and network performance across graphs with different spectra.
- Graph convolution: Order-K filters aggregate information through K-hop neighborhoods, and stacking convolutional layers enlarges the receptive field by K hops per layer.
- Pooling: Global pooling uses 1-max and variance statistics to produce representations invariant to point ordering and rotations, while multi-resolution pooling locally subsamples scattered points and aggregates their clusters.
- Final architecture: The final model uses two 1000-filter, order-3 graph-convolutional layers, optional multi-resolution pooling to 55 points, a final global pooling step, and a softmax classifier.
4. PERFORMANCE EVALUATION
PointGCN is evaluated on ModelNet10 and ModelNet40 using point-cloud inputs and compared with methods based on voxels, images, and point sets. The evaluation also examines training stability and how max pooling selects information across graph-convolutional layers.
- Dataset description and preparation: PointGCN is evaluated on ModelNet10 and ModelNet40, using uniformly sampled and then farthest-subsampled point clouds.ModelNet10 has 4,899 models across 10 categories, while ModelNet40 has 12,311 models across 40 categories.
- Performance comparison: Mean instance accuracy and mean category accuracy are used because both datasets have unbalanced class distributions.The comparison includes voxel-, image-, and point-set-based methods.
- Performance comparison: PointGCN improves over PointNet and ECC on ModelNet10 and ModelNet40 in both reported performance metrics, while remaining below MVCNN.MVCNN uses ImageNet1K pretraining and views objects from 80 perspectives; PointGCN uses no pretraining and one input view.
- Training stability: PointGCN shows faster convergence and more reliable performance than PointNet across both pooling approaches during ModelNet40 training.The comparison concerns fluctuations associated with sensitivity to initial weights and is illustrated through five training processes and 50-trial stability comparisons.
- Max-pooling visualization: Max pooling selects active points whose emphasized structure shifts from local patterns in the first layer to more global structure in the second.This visualization is consistent with localized graph convolutions exploring features from different receptive fields.
5. CONCLUSION
The paper concludes that PointGCN combines localized graph convolutions with point-cloud-specific global or multi-resolution pooling for 3D point-cloud classification. Its graph structure supports faster convergence, lower performance variance, order-invariant localized features, and robustness to rotation transformations.
- 5. CONCLUSION: PointGCN uses two fast localized graph-convolutional layers with global or multi-resolution pooling designed for point-cloud data.The approach is reported as competitive on ModelNet.
- 5. CONCLUSION: Graph geometry narrows the learned model’s search space, supporting faster convergence and improved robustness as measured by performance standard deviation.The conclusion attributes localized features to filters combining information from K-hop neighbors.
- 5. CONCLUSION: Symmetric graph-based filters preserve point-order invariance and are isotropic, making the learned model robust under rotation transformations.The paper connects this property to performance on ModelNet40 without using multiple input views.