Research map / 3D vision / Depth and stereo
Test3R: Learning to Reconstruct 3D at Test Time
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Test3R adapts DUSt3R at test time by tuning visual prompt tokens so that pointmaps of a shared reference view, predicted from two different image pairs, agree, improving 3D reconstruction and multi-view depth.
Problem
Pairwise pointmap regression in DUSt3R yields imprecise and mutually inconsistent geometry across image pairs, especially for short-baseline and out-of-distribution scenes, limiting global consistency of 3D reconstruction.
Approach
Using image triplets, predict reference-view pointmaps from two pairs and, at test time, optimize only learnable visual prompt tokens in the encoder with a self-supervised cross-pair consistency loss, requiring no labels.
Key results
- On 7Scenes, Test3R reaches mean Acc 0.105 and mean Comp 0.136 versus DUSt3R's 0.146 and 0.181.
- On multi-view depth, Test3R reports average rel 2.6 and tau 79.1 versus DUSt3R's 3.3 and 71.5.
- Applying Test3R to MASt3R and MonST3R improves their 7Scenes mean Acc and Comp (e.g. MonST3R Acc 0.240 to 0.218).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · 7Scenes | Acc mean: 0.146 -> 0.105 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · 7Scenes | Comp mean: 0.181 -> 0.136 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · 7Scenes | NC mean: 0.736 -> 0.746 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · NRGBD | Acc mean: 0.144 -> 0.083 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · NRGBD | Comp mean: 0.154 -> 0.079 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · NRGBD | NC mean: 0.871 -> 0.870 | Hurt |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · DTU | rel: 3.3 -> 2.0; reducing the Absolute Relative Error by 1.3 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · DTU | τ: 69.9 -> 84.1; increasing the Inlier Ratio by 14.2 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · ETH3D | rel: 3.3 -> 3.2 | Helped |
| add cross-pair consistency test-time learning with visual prompts | DUSt3R · ETH3D | τ: 73.0 -> 74.0 | Helped |
| add Test3R adaptation | MAST3R · 7Scenes | Comp mean: 0.211 -> 0.177 | Helped |
| add Test3R adaptation | MonST3R · 7Scenes | Comp mean: 0.268 -> 0.251 | Helped |
Surprising findings
- Test3R's multi-view depth results surpass several methods that use ground-truth poses and intrinsics, and some trained on the same domain.
- In the appendix, Test3R's median Acc and NC are not uniformly better than CUT3R and DUSt3R on NRGBD, and Test3R trails CUT3R on some 7Scenes metrics.
Limitations
- Final quality depends heavily on input images; struggles with occlusions, dynamic objects and illumination in the wild.
- Only reference-view pointmaps are used; source-view pointmaps and camera poses are not exploited.
- Triplet count grows as n^3 with many views, so view sampling is costly and remains open.
Experimental setup
Datasets: 7Scenes NRGBD DTU ETH3D
Benchmarks: 7Scenes NRGBD DTU ETH3D
Metrics: Acc Comp NC rel tau (inlier ratio at 3%)
Baselines: DUSt3R MASt3R MonST3R CUT3R Spann3R Fast3R VGGT RobustMVD COLMAP
Builds on: DUSt3R Visual Prompt Tuning Test-time training
Closest papers by meaning
- DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024 · similarity 0.90)
- MUSt3R: Multi-view Network for Stereo 3D Reconstruction (CVPR 2025 · similarity 0.89)
- Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors (CVPR 2025 · similarity 0.88)
- Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass (CVPR 2025 · similarity 0.88)
- MV-DUSt3R+: Single-Stage Scene Reconstruction from Sparse Views In 2 Seconds (CVPR 2025 · similarity 0.88)
- MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion (ICLR 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.