3D Shape Generation and Completion through Point-Voxel Diffusion
About
We propose a novel approach for probabilistic generative modeling of 3D shapes. Unlike most existing models that learn to deterministically translate a latent vector to a shape, our model, Point-Voxel Diffusion (PVD), is a unified, probabilistic formulation for unconditional shape generation and conditional, multi-modal shape completion. PVD marries denoising diffusion models with the hybrid, point-voxel representation of 3D shapes. It can be viewed as a series of denoising steps, reversing the diffusion process from observed point cloud data to Gaussian noise, and is trained by optimizing a variational lower bound to the (conditional) likelihood function. Experiments demonstrate that PVD is capable of synthesizing high-fidelity shapes, completing partial point clouds, and generating multiple completion results from single-view depth scans of real objects.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| 3D point cloud generation | ShapeNet Car (test) | 1-NNA (CD)64.49 | 57 | |
| 3D point cloud generation | ShapeNet Chair category (test) | MMD (CD)2.622 | 56 | |
| 3D point cloud generation | ShapeNet Airplane category (test) | 1-NNA (CD, %)73.82 | 55 | |
| Semantic Scene Completion | SemanticKITTI (test) | Overall IoU21.2 | 48 | |
| Point cloud generation | ShapeNet Car | 1-NNA (CD)54.55 | 27 | |
| Point cloud generation | ShapeNet chair | 1-NNA (CD)57.09 | 23 | |
| 3D Shape Generation | ShapeNet airplane | 1-NNA (CD)73.82 | 16 | |
| Point cloud generation | ShapeNet Chair (test) | 1-NNA (CD)54.6 | 16 | |
| 3D Scene Completion | SemanticKITTI (val) | JSD (BEV)0.498 | 14 | |
| Point cloud generation | ShapeNetPart Airplane (test) | 1-NNA (CD)73.8 | 13 |