Research map / 3D vision / Depth and stereo
LoMa: Local Feature Matching Revisited
David Nordström, Johan Edstedt, Georg Bökman, Jonathan Astermark, Anders Heyden, Viktor Larsson, Mårten Wadenbäck, Michael Felsberg, Fredrik Kahl
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ALIKED with DaD+DeDoDe and retrain on MegaDepth (ablation) | LightGlue-style matcher · HardMatch (validation) | mAA@10px: +5.7 | Improved |
| extend matcher and descriptor training data from MegaDepth to the full data mixture (ablation) | LoMa · HardMatch (validation) | mAA@10px: +9.1 | Improved |
| extend training from 50K to 250K steps (ablation) | LoMa · HardMatch (validation) | mAA@10px: +0.4 | Improved |
| scale matcher embedding dimension from 256 to 1024 (ablation) | LoMa · HardMatch (validation) | mAA@10px: +1.4 | Improved |
| use DaD detector rather than DISK for matching (ablation) | LoMa · HardMatch (validation) | mAA@10px: 45.3 -> 51.3 | Improved |
| specialize matcher on one detector instead of ensemble of detectors (ablation) | LoMa with DaD · HardMatch (validation) | mAA@10px: 51.0 -> 51.3 | Improved |
| increase keypoints from 512 to 4096 (ablation) | LoMa-B · MegaDepth-1500 | AUC@20: 80.3 -> 83.6 | Improved |
| increase keypoints from 2048 to 4096 (ablation) | SP+LG · MegaDepth-1500 | AUC@20: 76.4 -> 76.4 | No clear effect |
| train descriptor longer than 50K steps (ablation) | LoMa descriptor · HardMatch (validation) | validation loss | Hurt |
| replace ALIKED+LG with LoMa-G | sparse matcher · HardMatch (test) | mAA@10px: 35.7 -> 54.3 | Improved |
| replace ALIKED+LG with LoMa-G | sparse matcher · WxBS | mAA@10px: 43.9 -> 73.4 | Improved |
| replace ALIKED+LG with LoMa-G | sparse matcher · RUBIK | AUC@10deg: 49.0 -> 73.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- RoMa v2 — RoMa v2: Harder Better Faster Denser Feature Matching (ECCV 2026) · better on 2 of 4 shared results
- SP+SG — SPSG: Self-Supervised Photometric Scene Generation From RGB-D Scans (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Large Spatial Model: End-to-end Unposed Images to Semantic 3D (NeurIPS 2024 · similarity 0.86)
- LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation models (ICLR 2025 · similarity 0.86)
- JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba (CVPR 2025 · similarity 0.86)
- LoS: Local Structure-Guided Stereo Matching (CVPR 2024 · similarity 0.85)
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · similarity 0.85)
- GIM: Learning Generalizable Image Matcher From Internet Videos (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
