cs.CVMar 29, 2026
SaveMV-RoMa: From Pairwise Matching into Multi-View Track Reconstruction
Organizations: KAIST · Seoul National University
Abstract
Establishing consistent correspondences across images is essential for 3D vision tasks such as structure-from-motion (SfM), yet most existing matchers operate in a pairwise manner, often producing fragmented and geometrically inconsistent tracks when their predictions are chained across views. We propose MV-RoMa, a multi-view dense matching model that jointly estimates dense correspondences from a source image to multiple co-visible targets. Specifically, we design an efficient model architecture which avoids high computational cost of full cross-attention for multi-view feature interaction: (i) multi-view encoder that leverages pair-wise matching results as a geometric prior, and (ii) multi-view matching refiner that refines correspondences using pixel-wise attention. Additionally, we propose a post-processing strategy that integrates our model's consistent multi-view correspondences as high-quality tracks for SfM. Across diverse and challenging benchmarks, MV-RoMa produces more reliable correspondences and substantially denser, more accurate 3D reconstructions than existing sparse and dense matching methods. Project page: https://icetea-cv.github.io/mv-roma/.
Figures & tables
Figure 1: Overview of MV-RoMa. Given a source and multiple co-visible target images, MV-RoMa jointly estimates dense correspondence fields that are geometrically consistent across views (top). These fields are then fed into the SfM pipeline, finally yielding a dense and accurate reconstructed point cloud (bottom).
Figure 2: Pipeline of MV-RoMa. (a) Building Tracks. Given a source view and a set of target views , we first obtain initial pairwise matches from an off-the-shelf matcher and apply a sampling procedure to construct a sparse set of multi-view prior tracks . (b) Multi-view Track Prediction. The RGB images and prior tracks are fed into our Multi-view Encoder (Sec. 3.3 ), which uses track-based cross-attention to produce geometrically consistent dense features. The Multi-view Refiner (Sec. 3.4 ) then applies pixel-wise cross-attention to predict the final dense correspondences for all target views.
Figure 3: Pixel-aligned multi-view attention. Target features are warped to the source grid using , and per-pixel attention is performed across the aligned views. This avoids the quadratic cost of global cross-attention while refining fine-grained correspondence estimates (Sec. 3.4 ).
Figure 4: Confidence selection and reciprocity filtering. (a) For each pixel , we select the correspondence with the highest confidence from multiple predictions across groups. (b) We then apply bidirectional (forward–backward) consistency filtering, retaining matches whose cycle error is below the threshold (Sec. 3.5 ).
| Method | AUC - DLT | AUC - RANSAC |
| @1px / @3px / @5px | ||
| ALIKED [ 66 ] +LG [ 31 ] | 32.6 / 64.6 / 75.3 | 28.4 / 61.5 / 73.8 |
| LoFTR [ 49 ] | 30.2 / 53.3 / 61.6 | 38.0 / 66.9 / 76.2 |
| Mast3r [ 27 ] | 25.7 / 52.9 / 64.3 | 37.1 / 63.4 / 73.7 |
| UFM [ 10 ] | 27.4 / 58.8 / 70.8 | 29.2 / 61.3 / 73.2 |
| RoMa [ 13 ] | 41.0 / 67.9 / 76.9 | 44.7 / 72.6 / 81.4 |
Table 1: Homography estimation on HPatches. We report the area under the cumulative error curve (AUC) up to values of 1px, 3px and 5px, using DLT and RANSAC [ 15 ] as a homography solver.
| Method | Accuracy ( ) | Completeness ( ) | ||||
| 1cm | 2cm | 5cm | 1cm | 2cm | 5cm | |
| Type A. Including Two-view tracks | ||||||
| ALIKED + LG | 55.11 | 69.63 | 84.94 | 0.32 | 1.41 | 6.56 |
| LoFTR | 35.90 | 53.50 | 75.00 | 1.48 | 6.03 | 21.63 |
| Mast3R | 38.99 | 57.2 | 79.16 | 1.44 | 6.83 | 29.83 |
| UFM | 41.55 | 60.07 | 80.44 | 1.12 | 5.19 | 21.34 |
Table 2: Results of 3D Triangulation. Our method is compared with the baselines on the ETH3D [ 47 ] dataset using accuracy and completeness metrics with different thresholds.
| Type | Method | Texture-Poor SfM Dataset | IMC Dataset | ||||
| AUC@3 | AUC@5 | AUC@10 | AUC@3 | AUC@5 | AUC@10 | ||
| Detector-Based | COLMAP (SIFT+NN) [ 46 ] | 2.87 | 3.85 | 4.95 | 23.58 | 23.66 | 44.79 |
| SIFT [ 32 ] + NN + PixSfM [ 30 ] | 3.13 | 4.08 | 5.09 | 25.54 | 34.80 | 46.73 | |
| D2Net [ 12 ] + NN + PixSfM | 1.54 | 2.63 | 4.54 | 8.91 | 12.26 | 16.79 | |
| R2D2 [ 40 ] + NN + PixSfM | 3.79 | 5.51 | 7.84 | 31.41 | 41.80 | 54.65 | |
| SP [ 9 ] + SG [ 44 ] + PixSfM | 14.00 | 19.23 | 24.55 | 45.19 | 57.22 | 70.47 | |
Table 3: Multi-view camera pose estimation. AUC of pose error (%) at 3 ∘ , 5 ∘ , and 10 ∘ on the Texture-Poor SfM and IMC PhotoTourism datasets. Bold and underlined scores indicate the best and second-best results, respectively.
| Components | Accuracy | Completeness | |||
| MV-Encoder | MV-Refiner | 1cm | 2cm | 1cm | 2cm |
| 75.58 | 86.25 | 5.64 | 15.73 | ||
| ✓ | 77.80 | 87.38 | 6.75 | 17.04 | |
| ✓ | ✓ | 81.91 | 90.45 | 7.52 | 18.50 |
Table 4: Ablation on MV-Encoder and MV-Refiner. 3D triangulation results on ETH3D for model variants with different component combinations.
| Setup | AUC-DLT | ||
| @ 1px | @ 3px | @ 5px | |
| A. Number of Track Tokens (V=5) | |||
| 512 (Full) | 41.8 | 69.9 | 78.8 |
| 256 | 41.8 | 69.8 | 78.7 |
| 64 | 41.8 | 69.9 | 78.6 |
| 16 | 41.6 | 69.7 | 78.7 |
Table 5: Impact of multi-view priors and interactions. (A) The number of track tokens. (B) The number of interacting views.
Figure 5: Group Sampling Procedure. We use a two-stage strategy: Stage 1 greedily selects targets based on selection scores, while Stage 2 generates additional groups to enforce reciprocity.
| Method | Budget | Match | Select & Filter | COLMAP | Refine | Total |
| DF-SfM w/ RoMa | - | 1.5min | 2sec | 2.5min | 1.5min | 5.5min |
| MV-RoMa (Ours) | Full | 3min | 3sec | 3min | - | 6min |
| Half | 1.5min | 2sec | 2min | - | 3.5min |
Table 6: Runtime comparison of each method for SfM reconstruction on a 20-image scene from the Texture-Poor SfM dataset.
| Method | Budget | Texture-Poor AUC@3 ∘ / 5 ∘ / 10 ∘ |
| DF-SfM w/ RoMa | - | 45.22 / 63.54 / 80.54 |
| MV-RoMa (Ours) | Full | 60.71 / 74.66 / 85.88 |
| Half | 59.97 / 74.60 / 86.23 |
Table 7: Performance comparison between DF-SfM and MV-RoMa on the scene 1000 of the Texture-Poor SfM dataset.
| Matcher model | AUC @1px / @3px / @5px | |
| DLT | RANSAC | |
| ALIKED [ 66 ] + LightGlue [ 31 ] | 42.9 / 70.3 / 79.1 | 42.9 / 71.0 / 80.2 |
| UFM [ 10 ] | 41.8 / 69.9 / 78.8 | 42.6 / 71.5 / 81.0 |
Table 8: Homography estimation on the HPatches dataset using different off-the-shelf matchers to generate track tokens.
| Matcher model | Clustering | Error (median) | AUC @1px / @5px | |
| DLT | RANSAC | |||
| ALIKED [ 66 ] + LG [ 31 ] | 0.50 | 42.8 / 78.9 | 43.1 / 80.1 | |
| ✓ | 0.50 | 42.9 / 79.1 | 42.9 / 80.2 | |
| UFM [ 10 ] | 0.53 | 41.5 / 78.8 | 42.4 / 80.7 | |
| ✓ | 0.51 | 41.8 / 78.8 | 42.6 / 81.0 | |
Table 9: Homography estimation on the HPatches dataset, comparing MV-RoMa’s performance under clustering-based sampling versus random sampling.
Figure 6: Random vs Clustering. (a) Random sampling results in spatially uneven distributions with redundant aggregations (red circles). (b) Our clustering-based sampling approach generates a spatially uniform distribution of track tokens, ensuring better coverage for feature exchange.
| Border Masked | AUC @1px / @3px / @5px | |
| DLT | RANSAC | |
| 5% | 41.9 / 69.9 / 78.8 | 42.5 / 71.3 / 80.7 |
| 10% | 41.8 / 69.9 / 78.8 | 42.2 / 71.3 / 80.8 |
| 15% | 41.4 / 69.9 / 78.9 | 41.7 / 70.7 / 80.2 |
Table 10: Effect of reduced spatial coverage on homography estimation on HPatches.
Figure 7: Spatial Coverage Analysis. We mask the outer 5%, 10%, and 15% border regions (bright area) while maintaining a fixed total number of track tokens ( , only 300 tokens are shown for visualization). The experiment confirms that well-distributed tokens are essential for optimal results.
More in this topic
Image Matching
Image matching finds corresponding points, regions, or objects across images despite changes in viewpoint, scale, lighting, or sensor.
See all 84 papers