Source-linked AI summary
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
TL;DR
Dense feature matchers aim to estimate all image correspondences but still struggle with hard real-world pairs and can be computationally expensive. RoMa v2 combines a new matching architecture and objective, diverse training, efficient two-stage refinement, and DINOv3 features. It achieves state-of-the-art performance across diverse benchmarks while improving runtime over RoMa.
Problem
Existing dense matchers remain unreliable on difficult real-world scenarios, while high-precision models can be slow and memory-intensive.
Method
RoMa v2 combines a multi-view matching architecture and objective, diverse training data, a two-stage matcher-refiner pipeline, pixel-wise precision prediction, and frozen DINOv3 features.
Results
RoMa v2 sets a new state-of-the-art across diverse pose-estimation, localization, and dense-matching benchmarks while running 1.7× faster than RoMa.
Takeaways & Limitations
RoMa v2 matches or surpasses prior dense matchers across their respective strengths while producing harder, more precise matches at lower runtime than its predecessor.
Takeaways & Limitations
RoMa v2 is slightly less robust than RoMa to extreme modality changes, although it remains more robust than UFM in that setting.
Abstract
from arXiv · showhide
Dense feature matching aims to estimate all correspondences between two images of a 3D scene and has recently been established as the gold standard due to its high accuracy and robustness. However, existing dense matchers still fail or perform poorly for many hard real-world scenarios, and high-precision models are often slow, limiting their applicability. In this paper, we attack these weaknesses on a wide front through a series of systematic improvements that together yield a significantly better model. In particular, we construct a novel matching architecture and loss, which, combined with a curated diverse training distribution, enables our model to solve many complex matching tasks. We further make training faster through a decoupled two-stage matching-then-refinement pipeline, and at the same time, significantly reduce refinement memory usage through a custom CUDA kernel. Finally, we leverage the recent DINOv3 foundation model along with multiple other insights to make the model more robust and unbiased. In our extensive set of experiments, we show that the resulting novel matcher sets a new state-of-the-art, being significantly more accurate than its predecessors. Code is available at https://github.com/Parskatt/romav2
1 Introduction
Dense feature matching provides precise correspondences for downstream vision tasks, but existing matchers remain vulnerable to difficult scenarios and costly to run. RoMa v2 addresses these weaknesses with architectural, training, and efficiency improvements, achieving strong performance across benchmarks.
- Feature matching supports downstream tasks including visual localization and 3D reconstruction, which depend on precise, trustworthy correspondences.
- Dense matchers consider spatial context jointly across image pairs, extending beyond descriptor-only matching approaches.
- RoMa is robust to extreme appearance changes but remains challenged by extreme viewpoints and substantial runtime and memory demands.
- RoMa v2 achieves state-of-the-art results across a wide range of pose-estimation and dense-matching benchmarks.
- RoMa v2 combines a novel warp-and-correlation matching objective with faster, less memory-intensive refiners and diverse training data.
- The model adds pixel-wise error-covariance prediction for downstream refinement of estimated geometry.
2 Related Work
Recent feature matching has shifted from sparse descriptor matching toward learned, detector-free, and dense approaches. RoMa v2 builds on this progression with a two-stage architecture using foundation-model features, multi-view attention, and refined dense predictions.
- Feature Matching: Dense matching methods evolved from sparse keypoint pipelines toward detector-free and pixel-dense correspondence estimation.
- Feature Matching: RoMa improved robustness to extreme appearance changes by using frozen foundation-model features for coarse matching.
- RoMa v2 estimates coarse correspondences before refining them to sub-pixel accuracy in a two-stage pipeline.
- Its coarse matcher uses frozen DINOv3 features with a Multi-view Transformer, while refiners predict warp corrections, confidence changes, and precision matrices.
- Feed-forward Reconstruction: Feed-forward reconstruction methods increasingly incorporate matching while directly predicting scene structure or point maps from image pairs.
3 Method
RoMa v2 uses a two-stage coarse-matching and refinement architecture, combining frozen DINOv3 features, multi-view attention, new matching losses, efficient refiners, predictive covariance, and diverse training data. These changes target multi-view context, computational cost, subpixel accuracy, uncertainty estimation, and robustness across difficult matching scenarios.
- 3.1 Architecture: RoMa v2 trains matching and refinement separately: the frozen matcher first produces coarse warps, which the refiners then improve to subpixel accuracy.The refiners operate at strides 4, 2, and 1, reaching the original image resolution.
- 3.2 Matcher: The coarse matcher uses frozen DINOv3 features and a ViT-B Multi-view Transformer with alternating frame-wise and global attention.Dense Prediction Transformer heads output coarse warps and confidence at 4x downsampled resolution.
- 3.2 Matcher: RoMa v2 replaces RoMa’s single-view Gaussian Process matcher with multi-view attention and an auxiliary negative-log-likelihood target for patch correspondences.The matcher combines a robust warp-regression loss, overlap loss, and LNLL objective.
- 3.3 Refiners: A custom CUDA kernel reduces the memory use of local correlation in the refiners, addressing the high-resolution cost of RoMa’s implementation.The refinement design follows the coarse matcher’s stride-4 output and refines only at strides ≤4.
- 3.4 Data: A diverse mixture of wide- and small-baseline datasets improves robustness to large viewpoint changes, fine-grained details, and textureless surfaces.Aerial datasets support large rotations and air-to-ground views, while small-baseline data improves fine-grained prediction; EMA also substantially diminishes subpixel bias and improves accuracy.
4 Experiments
RoMa v2 achieves strong results across relative pose, visual localization, dense matching, and challenging geometric or multimodal benchmarks, while improving runtime over RoMa. Additional analyses examine robustness, backbone and refiner choices, covariance weighting, and the role of subpixel accuracy.
- 4.1 Relative Pose Estimation: RoMa v2 consistently outperforms prior matchers on MegaDepth-1500 and ScanNet-1500, surpassing reconstruction methods on MegaDepth and matching VGGT and MASt3R on ScanNet.These comparisons are reported in Table 4.
- 4.2 Visual Localization: RoMa v2 outperforms RoMa on Map-free and InLoc, reaching the top public-leaderboard result for InLoc.The InLoc result is described as a new state of the art.
- 4.3 Dense Matching: Across six dense-matching datasets, RoMa v2 outperforms previous methods, including UFM on TA-WB and RoMa on MegaDepth, with 84% lower EPE on AerialMegaDepth.The evaluation uses 640 × 640 inputs for RoMa v2 and RoMa, while UFM predictions are resized from its suggested inference resolution.
- 4.4 Ablations and Runtime Comparisons: RoMa v2 runs 1.7× faster than RoMa, is slightly slower than UFM, and uses a much smaller memory footprint than UFM.Table 8 benchmarks 640 × 640 images with batch size 8 on an H200.
- 4.5 Multi-Modal Matching on WxBS: On WxBS, RoMa v2 exceeds UFM but remains slightly below RoMa, while both RoMa v2 and UFM struggle on the IR-to-RGB multimodal subset.WxBS measures generalization under extreme viewpoint, illumination, and modality changes.
- 4.7 Matching across Geometric Challenges on RUBIK: RoMa v2 sets a new RUBIK state of the art with a 57.3 success ratio, exceeding DUSt3R’s 54.8.Success requires rotation error below 5° and translation error below 2 m.
5 Limitations and Future Work
RoMa v2 is slightly less robust than RoMa under extreme modality changes such as WxBS, although it remains significantly more robust than UFM.
- RoMa v2 is slightly less robust than RoMa under extreme modality changes, such as those evaluated by WxBS.The authors identify the trade-off between generalization and performance as future work.
6 Conclusion
RoMa v2 is a dense feature matcher designed for harder image pairs, more precise predictions, and faster runtime than its predecessor. These improvements produce denser matches by increasing the number of correct correspondences.
- RoMa v2 matches harder pairs with more precise predictions and faster runtime than RoMa, producing denser and more correct matches.
Matching Supplementary Material
The supplementary architecture details describe feature projection, multi-view Transformer processing, similarity construction, and efficient multi-scale refinement. The design uses DINOv3 features and refines predictions at progressively finer resolutions.
- Architecture: DINOv3 features from layers 11 and 17 are concatenated and projected from 2048 to 768 dimensions before multi-view Transformer processing.
- Architecture: The Transformer alternates global cross-frame attention with frame-wise attention using normalized Axial RoPE, producing 1024-dimensional outputs for both images.
- Matching: A temperature-scaled cosine-similarity matrix is constructed from the Transformer outputs and used to form match embeddings.
- Refinement: Refinement operates at strides 4, 2, and 1, with VGG19 features projected to 192, 48, and 12 channels at the corresponding resolutions.
- Refinement: Refiner inputs concatenate fine features, warped features, displacement features, and local correlations, with local-correlation neighborhoods shrinking from 7 to 3 to none across scales.
B Further Details on Datasets
The dataset supplement describes pair-construction procedures across diverse multi-view, synthetic, indoor, aerial, and optical-flow datasets. Sampling uses overlap thresholds, scene splits, and dataset-specific constraints.
- MegaDepth and AerialMegaDepth: MegaDepth and AerialMegaDepth sample up to 200,000 pairs per scene using directional-overlap thresholds above 0.01 and 0.35, with scene-size down-weighting.
- MapFree: MapFree training samples image pairs with overlap above 0.01 uniformly across scenes and uses only each scene’s seq0.
- ScanNet++ v2: ScanNet++ v2 uses 856 indoor training scenes, a minimum overlap of 0.2, and 10,000 qualifying pairs per scene, yielding approximately 8 · 10^6 pairs.
- TartanAir V2 and BlendedMVS: TartanAir V2 excludes five named scenes for testing, while BlendedMVS training uses pairs with directional overlap greater than 0.05.
- Hypersim: Hypersim trains on scenes with indices below 50 and validates on indices at least 50, using pairs with unidirectional overlap of at least 0.2.
C Further Details on Training Data
Training combines diverse, challenging image mixtures with light geometric and photometric augmentation, while covisibility is computed differently for depth- and flow-based datasets.
- Augmentations: Augmentations include aspect-ratio changes, flips, grayscale, brightness and hue jitter, plus random translations for MegaDepth and AerialMegaDepth.
- Training data: The training mixture spans diverse and challenging scene types, as illustrated by the sampled batches.
- Covisibility computation: Depth-based datasets define covisibility through depth consistency with threshold τ = 0.05.
- Covisibility computation: Flow-based datasets use warp cycle consistency below 5·10^-3, approximately 1.6 pixels at 640 × 640 resolution.
D Further Details on Benchmarks
SatAst is a satellite–astronaut matching benchmark built from correctly matching image pairs, with correspondences annotated iteratively and assessed through homography reprojection error.
- SatAst: SatAst contains 39 correctly matching astronaut–satellite image pairs selected from EarthMatch.
- Annotation: Corresponding points are annotated iteratively, using the procedure illustrated in Figure 9.
- Annotation: Annotation quality is evaluated by estimating homographies and measuring reprojection errors, whose distribution is plotted in Figure 10.
E Further Details on Predictive Covariance Experiment
The predictive covariance experiment evaluates match uncertainty on HyperSim and approximates full match covariance from RoMa v2’s forward and sampled backward covariances before robust geometric estimation.
- Benchmark: The benchmark uses 1,500 HyperSim validation-scene pairs after discarding pairs with overlap below 0.2.
- Covariance estimation: RoMa v2’s full 4 × 4 match covariance is approximated by a block-diagonal matrix combining forward and sampled backward covariances.
- Robust estimation: The experiment optimizes covariance-weighted Sampson error and uses the residual within MSAC scoring in standard LO-RANSAC.
- Baselines: VGGT evaluation samples 1,024 ALIKED keypoints and uses 0.1 confidence and covisibility thresholds, while UFM is evaluated at its fixed 420 × 560 resolution without upsampling.
G Bias In AerialMegaDepth
RoMa v2 can predict spurious overlap in textureless sky regions, which the authors associate with depth leakage from AerialMegaDepth’s scene meshes.
- Observed bias: RoMa v2 sometimes predicts overlaps in textureless sky regions, plausibly because AerialMegaDepth propagates mesh depths into sky pixels.
- Observed bias: The resulting sky pixels can appear multi-view consistent and may leak into the model’s warp estimates.