Introduction
Representing 3D point clouds with a fixed-length latent representation, or "codeword," is a fundamental challenge in 3D deep learning. Traditional autoencoders like FoldingNet or AtlasNet typically deform a continuous 2D square patch—a genus-zero primitive—into a 3D shape. While effective for single, solid objects, these methods struggle with complex topologies, such as objects with holes (e.g., a torus) or scenes containing multiple disjoint items. Because these models assume surface continuity, they often generate "ghost points"—noisy artifacts that bridge gaps between separate objects or stretch across holes, leading to poor latent representations that fail to capture the scene's true structure.
Methodology: The Tearing Module
To address these topological limitations, we introduce TearingNet. The core innovation is a decoder architecture that combines a traditional Folding module with a novel Tearing module.
The decoder consists of two primary components:
- Folding Network (F-Net): Maps 2D grid coordinates to 3D space based on the latent code.
- Tearing Network (T-Net): Takes the preliminary 3D reconstruction and "corrects" the 2D grid coordinates. By shifting points in the 2D domain, it can pull them far enough apart to effectively "tear" the underlying graph.
By treating the 2D grid as a local graph, TearingNet identifies edges between points that have been moved significantly apart and considers them "broken." This allows the model to represent holes or separate objects as disconnected sub-graphs, moving beyond the genus-zero constraint of previous models.
Experimental Setup
TearingNet was evaluated across several datasets designed to test topological complexity:
- ShapeNet: Standard single-object benchmark.
- Torus Dataset: Torus shapes with 1 to 3 holes.
- KIMO & CAMO (KITTI/CAD Multi-Object): Challenging datasets featuring scenes with multiple pedestrians, cars, and trucks synthesized from LiDAR scans and CAD models.
Results and Discussion
TearingNet demonstrates significant improvements in both reconstruction quality and downstream task performance.
Reconstruction Fidelity
Quantitatively, TearingNet outperforms FoldingNet and AtlasNet on multi-object datasets. Using the Earth Mover’s Distance (EMD), which is sensitive to point distribution, TearingNet achieved an error rate of 1.872 compared to FoldingNet's 3.056 on the KIMO-5 dataset. Qualitatively, TearingNet eliminates the "ghostly" connecting points that plague continuous-surface models.
Topology-Aware Latent Space
A key insight is that TearingNet's latent codewords are inherently topology-aware. A simple SVM classifier can accurately count the number of objects in a scene solely from the codeword. Furthermore, t-SNE visualizations reveal a tree-like organization of the latent space, where codewords for complex scenes (many objects) are recursively encapsulated by clusters representing simpler scenes.
Downstream Performance
The improved representations translate to better high-level understanding. On the KIMO-4 dataset, TearingNet reduced the Mean Absolute Error (MAE) in object counting by over 40% and improved accuracy in pedestrian detection tasks.
Conclusion and Future Work
TearingNet marks a significant step toward learning topology-friendly representations for 3D scenes. By explicitly modeling discontinuities, it provides a cleaner and more structured latent space. Future research will explore overcoming the bottleneck of fixed-length descriptors for increasingly complex scenes and applying the underlying Graph-Conditioned AutoEncoder (GCAE) framework to other domains, such as temporal topology in video processing.
Summary updated