This is a modified version of the original Hierarchical-Localization (hloc) toolbox, extended with state-of-the-art feature matching capabilities for improved visual localization performance.
We have integrated the following advanced matchers into the hloc pipeline:
- GIM-LightGlue matcher (standalone)
- RoMa + GIM-LightGlue matcher (RoMa for robust keypoint refinement + GIM-LightGlue matching)
These additions enable denser, more reliable matches and better 3D reconstructions, particularly in challenging scenes with low texture or repetitive patterns.
Explore the different configurations in the updated demo.ipynb notebook.
Uses GIM-LightGlue as the primary matching backbone.
| Stage | Results |
|---|---|
| Matching |    |
| Reconstruction |  |
Combines RoMa for enhanced keypoint detection/refinement with GIM-LightGlue matching.
| Stage | Results |
|---|---|
| Keypoints |   |
| Matching |    |
| Reconstruction |  |
-
GIM-LightGlue significantly increases the number of successful matches while preserving the original SuperPoint keypoints, resulting in denser correspondences and improved reconstruction quality.
-
RoMa + GIM-LightGlue delivers substantially superior performance: it discovers many additional matches from the original keypoints and generates a large number of entirely new matches thanks to the newly detected high-quality keypoints (particularly effective in low-texture). This leads to dramatically higher match counts, significantly more inliers, denser point clouds, and markedly more accurate 3D reconstructions.