Deep Unsupervised Learning of 3D Point Clouds via Graph Topology Inference and Filtering

Siheng Chen, Chaojing Duan, Yaoqing Yang, Duanshun Li, Chen Feng, Dong Tian

Abstract

We propose a deep autoencoder with graph topology inference and filtering to achieve compact representations of unorganized 3D point clouds in an unsupervised manner. Many previous works discretize 3D points to voxels and then use lattice-based methods to process and learn 3D spatial information; however, this leads to inevitable discretization errors. In this work, we handle raw 3D points without such compromise. The proposed networks follow the autoencoder framework with a focus on designing the decoder. The encoder adopts similar architectures as in PointNet. The decoder involves three novel modules: the folding module, the graph-topology-inference module, and the graph-filtering module. The proposed decoder leverages a learnable graph topology to push the codeword to preserve representative features and further improve the unsupervised-learning performance. Experimental results show that the proposed networks outperform state-of-the-art methods in reconstruction, visualization, and transfer classification.

Introduction

3D point clouds are the "raw" data produced by LiDAR and depth cameras, but they are notoriously difficult for AI to process because they are unorganized, irregular, and sparse. Most existing methods attempt to solve this by either voxelization—which causes massive memory overhead and quantization errors—or supervised learning, which requires expensive human-labeled data. This paper introduces a novel deep autoencoder designed to learn compact, representative features of 3D point clouds without any human labels, handling raw points directly.

Methodology

The authors propose an encoder-decoder architecture where the innovation lies primarily in the decoder. The encoder uses a PointNet-like architecture to compress a point cloud into a low-dimensional latent code. The decoder consists of three distinct modules:

  1. Folding Module: Takes a standard 2D lattice and "folds" it into 3D space based on the latent code, producing a coarse reconstruction.
  2. Graph-Topology Inference Module: Learns the pairwise relationships (edges) between points to represent the underlying structure of the object.
  3. Graph-Filtering Module: Uses the learned graph to apply a low-pass filter to the coarse reconstruction, smoothing the surface and refining point positions to capture fine details.
Evolution of 3D point cloud reconstruction from a 2D lattice across training epochs, comparing coarse folding with refined graph filtering.

The model is trained using an Augmented Chamfer Distance, a robust loss function that measures how closely the reconstructed point cloud matches the original.

Key Insights

  • Graph Smoothness: The authors prove that "graph smoothness" (points being similar to their neighbors in a learned graph) is a more powerful prior for 3D surfaces than simple spatial smoothness. This allows the model to handle complex shapes with holes.
  • Overcoming Folding Limits: A single 2D sheet cannot easily fold into complex 3D objects with holes without tearing. The graph-topology module effectively allows the network to "cut and glue" the 2D lattice to fit complex topologies.
  • The Decoder as a Teacher: By making the decoder more sophisticated, the encoder is forced to learn much richer, more descriptive latent features.

Experimental Results

The model was evaluated on standard benchmarks like ShapeNet and ModelNet:

  • Reconstruction: The proposed method achieved lower reconstruction errors than AtlasNet and 3DCapsNet across almost all categories in ShapeNetCore.
  • Transfer Classification: Features learned unsupervised on ShapeNet achieved 89.67% accuracy on ModelNet40, significantly outperforming Latent-GAN (85.7%).
  • Detail Preservation: Qualitative results demonstrate that graph filtering successfully recovers fine details, such as airplane engines, that are lost in models using only a folding module.

Conclusion and Future Work

This research demonstrates that a graph topology can be inferred as auxiliary information without specific supervision, leading to better performance in 3D point cloud tasks. While the $O(M^2)$ memory complexity of the graph adjacency matrix presents a scaling challenge, future work may explore divide-and-conquer strategies to handle higher-resolution point clouds.

Summary updated

Abstract

We propose a deep autoencoder with graph topology inference and filtering to achieve compact representations of unorganized 3D point clouds in an unsupervised manner. Many previous works discretize 3D points to voxels and then use lattice-based methods to process and learn 3D spatial information; however, this leads to inevitable discretization errors. In this work, we handle raw 3D points without such compromise. The proposed networks follow the autoencoder framework with a focus on designing the decoder. The encoder adopts similar architectures as in PointNet. The decoder involves three novel modules. The folding module folds a canonical 2D lattice to the underlying surface of a 3D point cloud, achieving coarse reconstruction; the graph-topology-inference module learns a graph topology to represent pairwise relationships between 3D points, pushing the latent code to preserve both coordinates and pairwise relationships of points in 3D point clouds; and the graph-filtering module couples the above two modules, refining the coarse reconstruction through a learnt graph topology to obtain the final reconstruction. The proposed decoder leverages a learnable graph topology to push the codeword to preserve representative features and further improve the unsupervised-learning performance. We further provide theoretical analyses of the proposed architecture. In the experiments, we validate the proposed networks in three tasks, including 3D point cloud reconstruction, visualization, and transfer classification. The experimental results show that (1) the proposed networks outperform the state-of-the-art methods in various tasks; (2) a graph topology can be inferred as auxiliary information without specific supervision on graph topology inference; and (3) graph filtering refines the reconstruction, leading to better performances.
BibTeX
@article{chen2020deep,
  author = {Siheng Chen and Chaojing Duan and Yaoqing Yang and Duanshun Li and Chen Feng and Dong Tian},
  title = {{Deep Unsupervised Learning of 3D Point Clouds via Graph Topology Inference and Filtering}},
  journal = {IEEE Trans. Image Process.},
  year = {2020},
  url = {https://openreview.net/forum?id=97RK3fjNrz8},
}
Figure