Research map / 3D vision / Depth and stereo
MAD-DR: Map Compression for Visual Localization with Matchness Aware Descriptor Dimension Reduction
Qiang Wang
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 |
|---|---|---|---|
| jointly retrain the matcher with compressed descriptors instead of using pre-trained SuperGlue (ablation) | MAD-DR (SuperPoint, D_LR=8) · Aachen Night (0.25m,2deg) | localization accuracy (%): 71.4 -> 82.7 | Improved |
| use asymmetric design (uncompressed query, compressed database) instead of symmetric training (ablation) | MAD-DR · InLoc duc1 / duc2 | localization accuracy (%): duc1 30.8 -> 46.5; duc2 29.0 -> 48.1 (vs Sym db only) | Improved |
| add descriptor reconstruction loss (ablation) | MAD-DR · InLoc duc1 | localization accuracy (%): 40.9 -> 46.5 | Improved |
| add descriptor reconstruction loss (ablation) | MAD-DR · Aachen Night | localization accuracy (%): 83.7 -> 82.7 | No clear effect |
| apply encoder-decoder to both query and database after asymmetric training (ablation) | MAD-DR · Aachen Night | localization accuracy (%): 82.7 -> 83.7 | No clear effect |
| compress SuperPoint descriptors to dimension 4 (1/256 size) with joint training | SuperPoint+SuperGlue · InLoc duc1 (0.25m,2deg) | localization accuracy (%): 49.5 -> 48.0 | No clear effect |
| joint training compression versus SceneSqueezer on Aachen night | SceneSqueezer · Aachen Night (0.25m,2deg) | localization accuracy (%): 50.0 -> 80.6 at 31 MB -> 4.1 MB | Improved |
| replace DISK+MNN+PQ baseline with joint-trained compression on InLoc | DISK+MNN+PQ · InLoc | localization accuracy (%): DISK+MNN does not perform well (>15% worse) on InLoc | Improved |
| apply method to DISK instead of product quantization | DISK · Aachen Night | localization accuracy (%): 70.4 -> 84.7 with 1/4 memory | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SceneSqueezer — SceneSqueezer: Learning To Compress Scene for Camera Relocalization (CVPR 2022) · better on 5 of 5 shared results
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
- Simple but Effective Triplet-Based Compression Strategies for Compact Visual Localization (CVPR 2026 · similarity 0.89)
- Hybrid Scene Compression for Visual Localization (CVPR 2019 · similarity 0.88)
- R-SCoRe: Revisiting Scene Coordinate Regression for Robust Large-Scale Visual Localization (CVPR 2025 · similarity 0.86)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.85)
- Fully Geometric Panoramic Localization (CVPR 2024 · similarity 0.85)
- RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations (CVPR 2025 · 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.
