Research map / 3D vision / Depth and stereo
RoMa v2: Harder Better Faster Denser Feature Matching
Johan Edstedt, David Nordström, Yushan Zhang, Georg Bökman, Jonathan Astermark, Viktor Larsson, Anders Heyden, Fredrik Kahl, Mårten Wadenbäck, Michael Felsberg
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 |
|---|---|---|---|
| add refiner EMA (decay 0.999) (ablation) | RoMa v2 refiners · MegaDepth-1500 | AUC@5deg: 61.4 -> 62.8 | Improved |
| add refiner EMA (ablation) | RoMa v2 refiners · ScanNet-1500 | AUC@5deg: 33.6 -> 33.6 | No clear effect |
| use DINOv3 instead of DINOv2 in coarse matcher (ablation) | RoMa v2 coarse matcher · WxBS | score: 35.6 -> 34.2 | Mixed |
| use DINOv3 instead of DINOv2 features (linear probe) (ablation) | frozen encoder · MegaDepth linear probe | EPE: 27.1 -> 19.0 | Improved |
| replace UFM matching objective with RoMa v2 attention plus NLL objective (ablation) | coarse matcher · Hypersim PCK@1px | PCK: 11.2 -> 30.5 | Improved |
| replace UFM training/architecture setup with RoMa v2 (ablation) | matcher · Hypersim PCK@5px | PCK: 67.4 -> 86.7 | Improved |
| use custom CUDA local correlation kernel (ablation) | RoMa v2 refiners · runtime | memory (GB): 5.6 -> 4.8 | Improved |
| use CUDA kernel for local correlation (ablation) | RoMa v2 refiners · runtime | throughput (pairs/s): 30.3 -> 30.9 | No clear effect |
| simulate less precise subpixel accuracy by perturbing predictions (ablation) | RoMa v2 · MegaDepth-1500 | AUC@5deg: 62.8 -> 46.3 | Hurt |
| upgrade RoMa to RoMa v2 (new matcher, data, refiners) | RoMa · Map-free | Precision: 59.7 -> 73.3 | Improved |
| upgrade RoMa to RoMa v2 | RoMa · AerialMegaDepth | EPE: 25.05 -> 4.12 | Improved |
| upgrade RoMa to RoMa v2 | RoMa · WxBS | mAA@10px: 60.8 -> 55.4 | Hurt |
Later papers that compared with this method
1 later accepted paper reports RoMa v2 as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- LoMa: Local Feature Matching Revisited (ECCV 2026 · better on 2 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
- RoMa: Robust Dense Feature Matching (CVPR 2024 · similarity 0.94)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.88)
- Diffusion Model for Dense Matching (ICLR 2024 · similarity 0.87)
- MATCHA: Towards Matching Anything (CVPR 2025 · similarity 0.87)
- Dense Match Summarization for Faster Two-view Estimation (CVPR 2025 · similarity 0.87)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 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.
