Research map / 3D vision / Depth and stereo
TTT3R: 3D Reconstruction as Test-Time Training
Xingyu Chen, Yue Chen, Yuliang Xiu, Andreas Geiger, Anpei Chen
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 |
|---|---|---|---|
| replace CUT3R's constant learning rate with confidence-guided per-token learning rate (TTT3R) | CUT3R · TUM-dynamic pose, 1000 frames | ATE: 0.173 -> 0.106 | Improved |
| replace CUT3R's constant learning rate with TTT3R rule | CUT3R · KITTI depth, 500 frames | delta<1.25: 80.2 -> 86.9 | Improved |
| use learned gating (ScalarLR/ConditionLR/TokenLR) instead of training-free closed-form rate (ablation) | CUT3R · TUM-dynamic pose, 1000 frames | ATE: 0.173 -> 0.154 (TokenLR best) vs 0.106 TTT3R | Improved |
| learned ScalarLR gating (ablation) | CUT3R · TUM-dynamic pose | RPE rot: 0.494 -> 0.502 | No clear effect |
| finetune with TTT3R update rule on 4-64 views (ablation) | TTT3R · TUM-dynamic pose | ATE: 0.106 -> 0.091 | Improved |
| finetune with TTT3R update rule on 4-64 views (ablation) | TTT3R · KITTI depth | Abs Rel: 0.131 -> 0.133 | Hurt |
| add periodic state reset every 100 frames to TTT3R (ablation) | TTT3R · TUM-dynamic pose, 1000 frames | ATE: 0.106 -> 0.093 | Improved |
| TTT3R vs reset/EMA/burn-in non-TTT baselines (ablation) | CUT3R · TUM-dynamic pose, 1000 frames | ATE: 0.126 (Reset) -> 0.106 | Improved |
| EMA shrinkage to initial state (ablation) | CUT3R · TUM-dynamic pose | ATE: 0.173 -> 0.164 | No clear effect |
| vary reset period 50/100/150 (ablation) | CUT3R + Reset · TUM-dynamic pose | ATE: 0.129 / 0.126 / 0.145 | Mixed |
| TTT3R on short sequences vs CUT3R | CUT3R · ScanNet 90 frames pose | ATE: 0.099 -> 0.064 | Improved |
| TTT3R vs offline VGGT | TTT3R · ScanNet 90 frames pose | ATE: 0.035 (VGGT) vs 0.064 | Hurt |
Later papers that compared with this method
10 later accepted papers report TTT3R as a baseline in their results tables; 10 of them report a better number on most of the results they share with it.
- VGG-T$^3$: Offline Feed-Forward 3D Reconstruction at Scale (CVPR 2026 · better on 10 of 12 shared results)
- RayMap3R: Inference-Time RayMap for Dynamic 3D Reconstruction (ECCV 2026 · better on 10 of 10 shared results)
- KV-Tracker: Real-Time Pose Tracking with Transformers (CVPR 2026 · better on 5 of 5 shared results)
- Stabilizing Streaming Video Geometry via Dynamic Feature Normalization (CVPR 2026 · better on 4 of 4 shared results)
- Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction (ECCV 2026 · better on 4 of 4 shared results)
- DROID-SLAM in the Wild (CVPR 2026 · better on 3 of 3 shared results)
- LASER: Layer-wise Scale Alignment for Training-Free Streaming 4D Reconstruction (CVPR 2026 · better on 2 of 2 shared results)
- Geometric Context Transformer for Streaming 3D Reconstruction (ECCV 2026 · better on 2 of 2 shared results)
2 more not listed.
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
- Test3R: Learning to Reconstruct 3D at Test Time (NeurIPS 2025 · similarity 0.90)
- tttLRM: Test-Time Training for Long Context and Autoregressive 3D Reconstruction (CVPR 2026 · similarity 0.90)
- Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction (CVPR 2026 · similarity 0.89)
- Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · similarity 0.89)
- Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors (CVPR 2025 · similarity 0.88)
- BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates (NeurIPS 2026 · 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.
