Research map / Label-efficient and robust learning / Recognition and long-tail
RoMa: Robust Dense Feature Matching
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.
Proposes RoMa, a robust dense feature matcher that combines frozen DINOv2 coarse features with ConvNet fine features, a transformer match decoder predicting anchor probabilities, and a regression-by-classification plus robust-regression loss.
Problem
Dense feature matching between two images must remain accurate under large real-world changes in scale, illumination, viewpoint and texture, but models trained from scratch or fine-tuned on limited 3D data overfit and generalize poorly.
Approach
Freeze DINOv2 as the coarse feature encoder for robustness, decouple it from a specialized ConvNet fine encoder for precise localization, decode coarse matches with a coordinate-agnostic transformer that outputs anchor probabilities, and use regression-by-classification for coarse loss and a generalized Charbonnier robust regression for refinement.
Key results
- Achieves 80.1 mAA@10px on WxBS versus 58.9 for DKM, a 36% improvement.
- Reaches 88.0 mAA@10 on IMC2022 versus 83.1 for DKM, and 62.6 AUC@5 on MegaDepth-1500 versus 60.4 for DKM.
- Ablation shows DINOv2 coarse features and the regression-by-classification plus robust loss each substantially lower 100-PCK@1px (14.4 to 13.1 across setups).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RoMa instead of the strongest listed baseline | Dense feature matching · WxBS | mAA@10px: 58.9 -> 80.1 | Helped |
| Replace shared coarse/fine encoder weights with separate encoders (ablation) | DKM-style matcher · Ablation validation | 100-PCK@1px: 17.0 -> 16.0 | Helped |
| Use VGG19 rather than RN50 for fine features (ablation) | Separate-encoder matcher · Ablation validation | 100-PCK@1px: 16.0 -> 14.5 | Helped |
| Use a Transformer rather than ConvNet match decoder (ablation) | VGG19-fine-feature matcher with baseline regression · Ablation validation | 100-PCK@1px: 14.5 -> 14.4 | Helped |
| Use DINOv2 coarse features (ablation) | Transformer-decoder matcher · Ablation validation | 100-PCK@1px: 14.4 -> 14.3 | Helped |
| Use regression-by-classification for coarse matching (ablation) | DINOv2 coarse-feature matcher · Ablation validation | 100-PCK@1px: 14.3 -> 13.6 | Helped |
| Use robust regression loss for refinement (ablation) | Regression-by-classification matcher · Ablation validation | 100-PCK@1px: 13.6 -> 13.1 | Helped |
| Replace the Transformer match decoder with a ConvNet (ablation) | RoMa · Ablation validation | 100-PCK@1px: 13.1 -> 14.0 | Hurt |
| Use frozen DINOv2 features rather than frozen RN50 features | Linear layer plus kernel nearest-neighbor matcher · MegaDepth validation | EPE: 60.2 -> 27.1 | Helped |
Surprising findings
- VGG19 features, which are worse than ResNet50 for coarse matching, outperform ResNet50 when used as fine features, indicating a tension between fine localizability and coarse robustness.
- Frozen ImageNet-style backbones (VGG19, ResNet50) give poor coarse matching out of the box, while frozen DINOv2 is far more robust (robustness 85.6% vs 43.2% for VGG19).
Limitations
- Gains are reported against a baseline (DKM) retrained by the authors; the frozen-backbone benefit is argued rather than exhaustively measured across backbones.
- Main model trained only on MegaDepth for most evaluations; ScanNet-specific model used for ScanNet-1500.
Experimental setup
Datasets: MegaDepth ScanNet
Benchmarks: WxBS IMC2022 MegaDepth-1500 ScanNet-1500 MegaDepth-8-Scenes InLoc
Metrics: mAA AUC@5/10/20 100-PCK EPE localization accuracy at 0.25/0.5/1.0 m
Baselines: DKM LoFTR PATS CasMTR PMatch ASpanFormer LightGlue DISK + LightGlue SuperPoint + SuperGlue QuadTree MatchFormer ASTR
Builds on: DKM DINOv2 LoFTR PDC-Net+
Closest papers by meaning
- MATCHA: Towards Matching Anything (CVPR 2025 · similarity 0.86)
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · similarity 0.85)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.85)
- SegMASt3R: Geometry Grounded Segment Matching (NeurIPS 2025 · similarity 0.84)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.84)
- XFeat: Accelerated Features for Lightweight Image Matching (CVPR 2024 · similarity 0.84)
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.