Research map / Point clouds / Shape and surface completion
Velox: Learning Representations of 4D Geometry and Appearance
Anagh Malik, Dorian Chan, Xiaoming Zhao, David B. Lindell, Oncel Tuzel, Jen-Hao Rick Chang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add texture augmentation (ablation) | Velox-S · Objaverse reconstruction (256 held-out scenes), small model | PSNR: 32.41 -> 33.93 | Improved |
| add texture augmentation (ablation) | Velox-S · Objaverse reconstruction (256 held-out scenes), small model | Chamfer (x1e5): 43.34 -> 38.79 | Improved |
| add texture augmentation (ablation) | Velox-S · Objaverse reconstruction (256 held-out scenes), small model | FVD: 99.16 -> 66.78 | Improved |
| encode jointly over time instead of per-frame LiTo | Velox · Objaverse reconstruction (256 held-out scenes) | PSNR: 32.55 -> 35.11 | Improved |
| use trained voxel decoder instead of sampled points for Gaussian decoder voxels (ablation) | Velox · Objaverse reconstruction (256 held-out scenes) | PSNR: 34.25 -> 35.11 | Improved |
| use decoded voxels instead of GT voxels (ablation) | Velox · Objaverse reconstruction (256 held-out scenes) | PSNR: 35.39 -> 35.11 | No clear effect |
| model time-conditioned geometry and appearance instead of deformations (GVF) | Velox · Objaverse reconstruction (256 held-out scenes) | PSNR: 26.45 -> 35.11 | Improved |
| use Heun 100 steps instead of Euler 10 steps at sampling (ablation) | Velox video-to-4D · Video-to-4D, Objaverse, conditioning view | PSNR: 22.19 -> 24.04 | Improved |
| use Velox tokens for video-to-4D instead of L4GM | Velox · Video-to-4D, Objaverse, novel view | PSNR: 18.30 -> 20.62 | Improved |
| use Velox tokens for tracking instead of CoTracker3 w/ GT depth | Velox · Objaverse 3D tracking | APD 3D (all points): 0.755 -> 0.835 | Improved |
| use Velox tokens for tracking instead of SpatialTrackerV2 | Velox · Objaverse 3D tracking | Occlusion accuracy: 0.897 -> 0.871 | Hurt |
Closest papers by meaning
- Learning Parallel Dense Correspondence From Spatio-Temporal Descriptors for Efficient and Robust 4D Reconstruction (CVPR 2021 · similarity 0.87)
- LiTo: Surface Light Field Tokenization (ICLR 2026 · similarity 0.87)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.85)
- Object-X: Learning to Reconstruct Multi-Modal 3D Object Representations (NeurIPS 2025 · similarity 0.85)
- Motion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking (CVPR 2024 · similarity 0.85)
- Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
