CoReMatch: Recurrent Continuous Coordinate Refinement for Real-Time Robotic Feature Matching
Organizations: Teledyne Technologies · Western University · Vector Institute
Abstract
Feature matching underpins many robotic perception tasks, including visual localization, mapping, SLAM, and autonomous navigation. Existing semi-dense matchers predict confidence on fixed grids, limiting correspondences to predefined locations and requiring costly high-resolution grids for precise localization. We present CoReMatch (Coordinate Refinement Match), a recurrent framework that refines correspondences directly in continuous coordinate space rather than predicting discrete confidence maps. CoReMatch progressively improves correspondence estimates by propagating information across iterations. At the coarse stage, a Deep Aggregated Transformer captures global context and iteratively updates correspondence coordinates, while a lightweight regression head further sharpens their localization. Experiments on indoor and outdoor benchmarks show that CoReMatch strikes a strong balance between accuracy and efficiency, outperforming Efficient LoFTR in accuracy while running more than 10x faster on embedded edge hardware. These results demonstrate that recurrent coordinate refinement offers an efficient and scalable approach to real-time feature matching for robotic perception.
Figures & tables
| Category | Method | MegaDepth | ScanNet | ||
| AUC@ / / | Time | AUC@ / / | Time | ||
| Sparse | SP [ 6 ] + SG [ 22 ] | 49.7 / 67.1 / 80.6 | - | 16.2 / 32.8 / 49.7 | - |
| SP [ 6 ] + LG [ 15 ] | 49.9 / 67.0 / 80.1 | 548 | 14.8 / 30.8 / 47.5 | 461 | |
| Dense | DKM [ 8 ] | 60.4 / 74.9 / 85.1 | 7110 | 26.6 / 47.1 / 64.2 | 1418 |
| ROMA [ 9 ] | 60.9 / 75.6 / 85.6 | 4964 | 28.9 / 50.4 / 68.3 | 4964 | |
| Semi-Dense | LoFTR [ 24 ] | 52.8 / 69.2 / 81.2 | 2465 | 16.9 / 33.6 / 50.6 | 437 |
| Dataset | LoFTR | ELoFTR | CoReMatch | vs. ELoFTR |
| TUM [ 23 ] | 0.5931 | 0.5932 | 0.5934 | -0.03% |
| EuRoC [ 3 ] | 2.3859 | 2.2769 | 2.0673 | +9.2% |
| KITTI [ 10 ] | 10.5482 | 10.1711 | 6.9788 | +31.4% |
| TartanAir [ 32 ] | 35.0914 | 35.4660 | 33.7973 | +4.7% |
| DAT | WAFT [ 34 ] | SCT | PTA [ 28 ] | |
| Reprojection AUC@3px (%) | 95.5 | 67.7 | 89.1 | 92.2 |
| Inference Time (ms) | 4.6 | 19.3 | 4.1 | 4.7 |
| Iteration | Top- Ratio (%) | |||||||
| 0 | 1 | 2 | 3 | 10 | 20 | 35 | 50 | |
| Reprojection AUC@3px (%) | 29.2 | 92.8 | 95.5 | 95.9 | – | – | – | – |
| Relative Pose AUC@10 (%) | – | 71.0 | 71.6 | 71.7 | 67.4 | 70.5 | 71.6 | 71.6 |
| Inference Time (ms) | – | 72.9 | 86.5 | 98.1 | 83.4 | 84.2 | 86.5 | 87.5 |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Resolution | Pose Est. AUC | Runtime (ms) | Memory (MB) | |
| @ /@ /@ | Orin | 3090 | |||
| ELoFTR | 512 512 | 45.7/62.9/76.1 | 191.4 | 11.7 | 822 |
| 640 640 | 51.0/67.4/79.8 | 310.6 | 18.2 | 932 | |
| 800 800 | 53.4/70.0/81.9 | 668.1 | 42.2 | 2652 | |
| 960 960 | 54.7/70.7/82.4 | 1394.9 | 48.7 | 3046 | |
| 1152 1152 | 56.4/72.2/83.5 | 2066.7 | 66.7 | 3574 | |