cs.CVOct 23, 2026
SaveFlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation
Organizations: ETH Zurich · Google · Microsoft Spatial AI Lab
Abstract
To estimate correspondences across images, existing approaches match either sparse interest points, which is highly efficient but limited by repeatability, or dense pixel grids, which is robust but expensive. Furthermore, these approaches are trained with only matching pairs and thus often fail to reject unrelated pairs that exhibit visual symmetries, which is equally important in applications like Structure-from-Motion (SfM). We bridge these gaps by introducing FlexGlue, a neural network for asymmetric sparse-to-dense matching and disambiguation. Given two images and coordinates of points in the first image, FlexGlue estimates the coordinates of the corresponding points in the second image while simultaneously rejecting the pair if non-matching. By searching for a limited set of correspondences among dense pixels, our approach is both fast and effective at both matching and disambiguation: it combines the sub-pixel accuracy of dense matching with the speed of sparse matching, while rejecting incorrect pairs with high precision. Experiments on multiple datasets show that FlexGlue pushes the Pareto frontier in relative pose estimation and significantly outperforms existing approaches for SfM and visual localization. FlexGlue is publicly available at https://github.com/cvg/flexglue .
Figures & tables
Figure 2: Architecture of FlexGlue. Each of the two images is encoded by a CNN backbone. The resulting dense tokens are fused by a Transformer that outputs context-augmented tokens. These tokens are decoded into overlap heatmaps and classify whether the image pair is unmatchable, in which case the inference stops. Otherwise, a query point in one image is turned into a sub-pixel coordinate in the other by interpolation of its feature, dense correlation, and match refinement.
Figure 3: Visual disambiguation examples. Left: Images overlaid with PCA of learned features. Center: Estimated correspondences pre-filtering. Right: We report the covisibility score after 3 and 9 layers in the top-left. Top: Our method correctly identifies a covisible pair. Bottom: We correctly identify challenging doppelgangers from various domains early and stop matching.
Figure 4: Hard negatives of our LaMAR doppelganger benchmark. From left to right: different floors (CAB), indoor (HGE), outdoor (HGE), and cross-device (LIN) pairs. None of the pairs overlap, despite a high retrieval similarity (sim) from MegaLoc; and denote the horizontal and vertical distance between the two cameras.
| Detector | Matching | MegaDepth-1500 | ScanNet-1500 | HardMatch | Extr. | Match | ||||||
| mAA@10px | [ms] | [ms] | ||||||||||
| ALIKED | LightGlue | 52.4 | 68.8 | 81.0 | 17.0 | 34.2 | 51.3 | 36.3 | 36.0 | 21.7 | ||
| DaD | LoMa-G | 56.1 | 72.2 | 84.0 | 29.3 | 51.7 | 70.0 | 54.3 | 128.5 | 49.8 | ||
| Sparse | 4096 kp. | SuperPoint | Ours | 53.8 | 70.0 | 82.0 | 31.9 | 54.8 | 72.6 | 55.7 | 33.1 | 29.6 |
| - | RoMa | 62.6 | 76.7 | 86.3 | 31.8 | 53.4 | 70.9 | 48.1 | 49.4 | 239.5 | ||
| - | MASt3R | 54.5 | 71.9 | 84.0 | 34.0 | 57.0 | 74.4 | 33.5 | 51.1 | 747.2 | ||
Table 1: Relative pose estimation on MegaDepth-1500, ScanNet-1500 and HardMatch. We report the AUC of the pose error at , , and , the HardMatch mAA@10px, and the time for extraction and matching. We highlight the first , second , and third best results. FlexGlue has comparable accuracy with dense matchers at a fraction of their cost, and also improves sparse-to-sparse feature matching on ScanNet and HardMatch.
| Ground–Ground | Ground–Aerial | Aerial–Aerial | ||||||||||
| Method | HPatches | WxBS | RUBIK | IMC22 | TerraSky3D | EarthMatch | SatAst | |||||
| 3px | 5px | 10px | mAA@10px | succ. | mAA | % loc. | 10px | |||||
| ALIKED+LG | 66.2 | 76.9 | 86.3 | 43.9 | 52.8 | 59.1 | 39.0 | 79.2 | 25.3 | 29.4 | 84.0 | 18.4 |
| LoMa-G | 66.4 | 77.5 | 87.3 | 73.4 | 73.1 | 79.8 | 47.6 | 89.8 | 53.4 | 61.6 | 98.0 | 24.4 |
| LoMa-R | 66.8 | 77.9 | 87.7 | 65.5 | 68.6 | 75.9 | 45.1 | 86.3 | 44.2 | 51.7 | 99.6 | 48.3 |
| RoMav2 | 72.2 | 81.2 | 89.2 | 64.8 | 78.4 | 86.2 | 59.5 | 84.7 | 53.9 | 62.9 | 92.2 | 38.1 |
Table 2: Two-view matching in additional domains. FlexGlue exhibits the highest matching accuracy over multiple datasets from different domains. A higher resolution further improves results.
| Method | Labelled data | DG | ViSym-F | LaMAR | Time | |||
| AUROC | AP | AUROC | AP | AUROC | AP | [ms] | ||
| DG ( Cai et al., 2023 ) | DG | 94.4 | 95.4 | 66.2 | 74.1 | 69.5 | 69.8 | 139.5 |
| DG++ ( Xiangli et al., 2025 ) | DG | 98.1 | 98.0 | 86.4 | 90.9 | 78.6 | 79.4 | 124.7 |
| DG++ ( Xiangli et al., 2025 ) | DG + ViSym | 98.1 | 98.1 | 88.3 | 90.7 | 73.3 | 75.3 | 124.7 |
| Ours | matching | 97.0 | 97.1 | 95.3 | 96.5 | 88.5 | 89.0 | 28.7 |
| Ours | DG + matching | 98.4 | 98.2 | 95.9 | 96.8 | 88.8 | 89.2 | 28.7 |
Table 3: Detection of Doppelgangers. We report AUROC and AP on DG (in-distribution), ViSym and LaMAR, as well as the matching time. FlexGlue achieves the highest disambiguation accuracy across all datasets when trained on the same labelled data (DG), especially +10 AUROC points on the LaMAR benchmark, at much higher throughput.
Table 8: Doppelgangers ablation. AUROC on ViSym-F and LaMAR Disambiguation. Our model learns disambiguation early, and guidance from match supervision improves generalization. Even without labelled data our model achieves high accuracy. Training in-domain increases accuracy. Right: Qualitative examples with GT, DG++ ( Xiangli et al., 2025 ) and our disambiguation score.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Run-time on different keypoint budgets and layer ablations . RoMav2’s ( Edstedt et al., 2026 ) run time is independent of the number of keypoints, but has a large overhead. LoMa ( Nordström et al., 2026 ) is very efficient but scales quadratically, while our method is fairly efficient and scales linearly (left). Our method learns strong matching and disambiguation early on, while the run time grows linearly with the number of layers (right, black line).
| Keypoints | ETH3D SfM | IMC2025 | HardMatch | |||
| score | mAA | |||||
| SuperPoint | 68.9 | 86.8 | 92.3 | 95.3 | 66.6 | 55.5 |
| DaD | 70.1 | 86.9 | 91.0 | 94.5 | 65.7 | 55.3 |
| RaCo | 71.7 | 88.1 | 92.5 | 95.4 | 66.6 | 54.9 |
Table 9: Keypoint detectors. FlexGlue works with any detector. RaCo yields the most accurate reconstructions, while all detectors perform similarly on two-view matching.
Figure 6: Detailed results on LaMAR-DG. AUROC and AP of FlexGlue and DG++ ( Xiangli et al., 2025 ) per scene (a), difficulty (b), time (c), and device (d). Difficulty bins are tertiles of MegaLoc similarity per label, i.e . , hard negatives are highly similar and hard positives dissimilar. FlexGlue remains robust on hard, cross-day, and cross-device pairs, where DG++ degrades strongly.
| Original label | Pairs | Unchanged | Flipped | Unsure |
| Positive | 1590 | 1500 | 57 | 33 |
| Negative | 1590 | 1154 | 418 | 18 |
| Total | 3180 | 2654 | 475 | 51 |
| Kept | 1184 | 946 | 238 | 0 |
Table 10: ViSym test relabelling. We manually clean GPS-annotated labels of ViSym by hand. Most corrections turn negatives into positives. The last row counts the pairs kept in our balanced subset.
Figure 7: Corrected ViSym test labels. Each pair shows the original ViSym label (Orig.) and our label (New). The original labels miss covisible pairs under strong viewpoint changes and mark some pairs of distinct but similar facades as covisible.
| Dataset | Type | #Scenes | Weight |
| MegaDepth ( Li & Snavely, 2018 ) | outdoor, real | 152 | 0.17 |
| AerialMegaDepth ( Vuong et al., 2025 ) | aerial–ground, rendered | 113 | 0.17 |
| BlendedMVS ( Yao et al., 2020 ) | objects & scenes, rendered | 493 | 0.15 |
| ScanNet++ ( Yeshwanth et al., 2023 ) | indoor, real | 827 | 0.15 |
| Hypersim ( Roberts et al., 2021 ) | indoor, synthetic | 399 | 0.10 |
| Map-free ( Arnold et al., 2022 ) | outdoor, real | 460 | 0.10 |
Table 11: Training datasets. Number of training scenes and relative sampling weight of each dataset. Doppelgangers is only mixed in during fine-tuning with negatives.
Figure 8: Comparison of matching modes. Left: Point-to-point correspondences estimated by our model, colored by epipolar error (green px). Center: Keypoints colored by being matchable sparse-to-sparse (green), and keypoints only matchable sparse-to-dense (blue), highlighting the repeatability issue. Right: Sparse-to-dense correspondences. Especially under large scale changes and on low-texture areas, sparse-to-dense matching yields many more, accurate correspondences.
Figure 9: PCA of the aggregated tokens. For each image pair, we show the two views and the first three principal components of their tokens at 9 layers as RGB. Our method learns structural priors such as horizontal and vertical lines, perspective, and planarity, and clearly highlights covisible areas.
Figure 10: Scene clustering on IMC2025. Image graphs recovered by our method with visual disambiguation only. Colors denote the predicted clusters and edges the retained image pairs. Top: landmarks are separated into clean clusters. Bottom: a single wrong pair (red) passes our filter.
Figure 11: Dense matching. Dense correspondences of FlexGlue, colored by their error from green ( px) to red ( px), with the PCK@1px per image, from the coarse readout, with our refinement, and with the RoMav2 ( Edstedt et al., 2026 ) refiner. The coarse matching has weaknesses near depth discontinuities, which the refiners address.
Figure 12: Dense matching across layers. Unfiltered dense correspondences of FlexGlue after 3, 6, and 9 layers, colored by their error from green ( px) to red ( px), with the PCK@1px per image. Despite being supervised only at sparse queries, FlexGlue yields fairly accurate dense correspondences. Deeper layers mostly refine matches and fix errors close to depth discontinuities.
Figure 13: Visual disambiguation examples. We show some visual disambiguation examples from the test sets of DG ( Cai et al., 2023 ) , ViSym-F ( Xiangli et al., 2025 ) , and LaMAR-DG ( Sarlin et al., 2022 ) . Top left is the GT label, right our covisibility score. Our method effectively distinguishes hard covisible pairs, but sometimes mislabels under extreme appearance or viewpoint changes, or when context is missing (middle row failure, wrong facade).
Figure 14: Qualitative matches. Correspondences of LoMa-G, SP+RoMav2, and SP+Ours on two pairs per dataset, colored by their reprojection error. We report the number of matches and the precision at 3 px. LoMa gets 2048 keypoints, and RoMav2 and our method get 1024 keypoints for fair comparison.
Figure 14: Qualitative matches (continued). Bottom: failure cases on TartanAir, Hypersim, and MegaDepth. All methods fail under extreme rotation of a repetitive ceiling (top) and on a strongly self-similar scene (middle). On the repetitive facade (bottom), our method matches the wrong instance of a repeated element, while RoMav2 succeeds.
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 87 papers