Research map / 3D vision / Depth and stereo
MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE
Ruijie Zhu, Jiahao Lu, Wenbo Hu, Xiaoguang Han, Jianfei Cai, Ying Shan, Chuanxia Zheng
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 |
|---|---|---|---|
| fine-tune the whole VAE with mean normalization instead of decoder-only with max normalization (ablation) | Geometry VAE · Sintel | Rel^p: 20.76 -> 5.68 | Improved |
| fine-tune the whole VAE with mean normalization instead of decoder-only with max normalization (ablation) | Geometry VAE · Monkaa | Rel^p: 14.44 -> 5.03 | Improved |
| use mean-rescaled VAE in the U-Net stage instead of max-rescaled (ablation) | Geometry U-Net · Sintel | Rel^p: 40.47 -> 35.17 | Improved |
| use mean-rescaled VAE in the U-Net stage instead of max-rescaled (ablation) | Geometry U-Net · Monkaa | Rel^p: 33.66 -> 27.36 | Improved |
| add depth-map supervision to Geometry VAE training (ablation) | Geometry VAE · ScanNet | Rel^p: 4.46 -> 3.03 | Improved |
| add depth-map supervision to Geometry VAE training (ablation) | Geometry VAE · Sintel | Rel^p: 5.68 -> 4.39 | Improved |
| add decoder loss when training the U-Net (ablation) | Geometry U-Net · DDAD | Rel^p: 33.78 -> 22.39 | Improved |
| add decoder loss when training the U-Net (ablation) | Geometry U-Net · Sintel | delta^p: 58.08 -> 56.94 | Hurt |
| use unified fusion of geometry and motion latents instead of separate VAEs (ablation) | Motion U-Net · Spring | EPE: 6.37 -> 5.16 | Improved |
| use unified fusion of geometry and motion latents instead of separate VAEs (ablation) | Motion U-Net · Point Odyssey | EPE: 3.65 -> 3.49 | Improved |
| use unified fusion instead of separate motion VAE (ablation) | Motion VAE · Spring | EPE: 0.66 -> 0.88 | Hurt |
| use deterministic paradigm instead of denoising diffusion paradigm (ablation) | MotionCrafter · Sintel | Rel^p: 35.95 -> 32.46 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 1 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GeometryCrafter: Consistent Geometry Estimation for Open-world Videos with Diffusion Priors (ICCV 2025 · similarity 0.90)
- Motion 3-to-4: 3D Motion Reconstruction for 4D Synthesis (CVPR 2026 · similarity 0.89)
- TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking (NeurIPS 2026 · similarity 0.88)
- Flux4D: Flow-based Unsupervised 4D Reconstruction (NeurIPS 2025 · similarity 0.87)
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.87)
- Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos (CVPR 2025 · similarity 0.87)
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.
