Research map / Video / Video segmentation and tracking
How Much 3D Do Video Foundation Models Encode?
Zixuan Huang, Xiang Li, Zhaoyang Lv, James M. Rehg
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 |
|---|---|---|---|
| use temporal video features instead of per-frame DINOv2 features (ablation) | 3D probe · DL3DV | point map error / AUC@30: Point 2.814 -> 1.576, AUC@30 0.245 -> 0.558 (DINOv2 vs V-JEPA) | Improved |
| use temporal video features instead of per-frame DINOv2 features (ablation) | 3D probe · CO3Dv2 | point map error / AUC@30: Point 0.559 -> 0.439, AUC@30 0.508 -> 0.619 (DINOv2 vs V-JEPA) | Improved |
| add 3D-aware finetuning (Aether) to CogVideoX | CogVideoX · DL3DV | point map error / AUC@30: Point 1.748 -> 1.566, Depth 0.608 -> 0.574, AUC@30 0.486 -> 0.527 | Improved |
| add 3D-aware finetuning (Aether) to CogVideoX | CogVideoX · CO3Dv2 | point and depth error: Point 0.485 -> 0.501, Depth 0.231 -> 0.249 | Hurt |
| scale WAN2.1 from 1.3B to 14B (ablation) | WAN2.1 · CO3Dv2 ablation subset | point map error: 0.0468 -> 0.0360 | Improved |
| scale CogVideoX from 2B to 5B (ablation) | CogVideoX · CO3Dv2 ablation subset | point map error: 0.0576 -> 0.0590 | Hurt |
| replace DINO features with frozen WAN2.1-14B features in VGGT | VGGT · CO3Dv2 | point map error / AUC@30: Point 0.476 -> 0.289, AUC@30 0.565 -> 0.718 | Improved |
| replace DINO features with frozen WAN2.1-14B features in VGGT | VGGT · DL3DV | point map error / AUC@30: Point 2.751 -> 1.034, AUC@30 0.363 -> 0.686 | Improved |
| halve probe width from 1024 to 512 (ablation) | 3D probe (WAN2.1-14B) · DL3DV | point map error: 1.051 -> 1.317 | No clear effect |
| use WAN2.1-14B features instead of Fast3R features | 3D probe · DL3DV | point map error / AUC@30: Point 1.379 -> 1.051, AUC@30 0.637 -> 0.660 | Improved |
| use WAN2.1-14B features instead of Fast3R features | 3D probe · CO3Dv2 | point map error / AUC@30: Point 0.262 -> 0.284, AUC@30 0.769 -> 0.736 | Hurt |
| extract features from mid-network layers at early-but-not-first timestep (ablation) | video diffusion models · CO3Dv2 ablation subset | point map error | Improved |
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.
- Aether — Aether: Geometric-Aware Unified World Modeling (ICCV 2025) · better on 4 of 4 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
- Probing the 3D Awareness of Visual Foundation Models (CVPR 2024 · similarity 0.91)
- Feat2GS: Probing Visual Foundation Models with Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Lexicon3D: Probing Visual Foundation Models for Complex 3D Scene Understanding (NeurIPS 2024 · similarity 0.88)
- Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video (CVPR 2025 · similarity 0.87)
- Can 4D Foundation Models Remember? (NeurIPS 2026 · similarity 0.87)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.86)
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.
