Research map / 3D vision / Depth and stereo
Geo4D: Leveraging Video Generators for Geometric 4D Scene Reconstruction
Zeren Jiang, Chuanxia Zheng, Iro Laina, Diane Larlus, Andrea Vedaldi
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 |
|---|---|---|---|
| train with disparity and ray maps in addition to point maps (ablation) | Geo4D · Sintel video depth | Abs Rel: 0.232 -> 0.223 | Improved |
| train with disparity and ray maps in addition to point maps (ablation) | Geo4D · Sintel camera pose | ATE: 0.335 -> 0.237 | Improved |
| fuse all three modalities at inference via multi-modal alignment vs disparity only (ablation) | Geo4D · Sintel video depth | Abs Rel: 0.211 -> 0.205 | Improved |
| fuse all three modalities at inference vs ray map only (ablation) | Geo4D · Sintel camera pose | ATE: 0.268 -> 0.185 | Improved |
| add point map to training and inference (ablation) | Geo4D · Sintel camera pose | ATE: 0.322 -> 0.185 | Improved |
| decrease sliding window stride from 15 to 4 (ablation) | Geo4D · Sintel video depth | Abs Rel: 0.213 -> 0.205 | Improved |
| decrease stride from 4 to 2 (ablation) | Geo4D · Sintel | s / frame: 1.89 -> 3.26 | Mixed |
| use fine-tuned point map VAE with uncertainty branch (ablation) | Geo4D · Sintel video depth | Abs Rel: 0.212 -> 0.205 | Improved |
| increase DDIM steps from 1 to 5 (ablation) | Geo4D · Sintel video depth | Abs Rel: 0.221 -> 0.205 | Improved |
| increase DDIM steps from 5 to 25 (ablation) | Geo4D · Sintel video depth | Abs Rel: 0.205 -> 0.220 | Hurt |
| Geo4D vs DepthCrafter on Sintel | Geo4D · Sintel video depth | Abs Rel: 0.270 -> 0.205 | Improved |
| Geo4D vs MonST3R camera translation on Sintel | Geo4D · Sintel camera pose | ATE: 0.108 -> 0.185 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MonST3R — MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion (ICLR 2025) · better on 8 of 12 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
- Syn4D: A Multiview Synthetic 4D Dataset (ECCV 2026 · similarity 0.91)
- Diffusion4D: Fast Spatial-temporal Consistent 4D generation via Video Diffusion Models (NeurIPS 2024 · similarity 0.90)
- CAT4D: Create Anything in 4D with Multi-View Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos (CVPR 2025 · similarity 0.89)
- IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer (NeurIPS 2026 · similarity 0.88)
- GeoVideo: Introducing Geometric Regularization into Video Generation Model (NeurIPS 2025 · similarity 0.88)
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.
