Research map / 3D vision / Depth and stereo
Investigating the Role of Image Retrieval for Visual Localization
Martin Humenberger, Yohann Cabon, Noé Pion, Philippe Weinzaepfel, Donghwan Lee, Nicolas Guérin, Torsten Sattler, Gabriela Csurka
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 |
|---|---|---|---|
| use cosine-similarity weights (CSI) instead of barycentric descriptor interpolation (BDI) | pose approximation · Aachen, RobotCar, Baidu, GangnamStation_B2 | % localized (5m, 10deg) | Improved |
| weight top-k poses by similarity (BDI or CSI) instead of equal weights (EWB) | pose approximation (Task 1) · multiple datasets | % localized (5m, 10deg) | Improved |
| increase number of retrieved images k beyond 3 | pose approximation (Task 1) · multiple datasets | % localized (5m, 10deg): k = 3 sweet spot; drops for k > 3 | Hurt |
| retrieve k>1 (up to k=4) images instead of the top-1 pose | pose approximation (Task 1) · Aachen | % localized (5m, 10deg) | Mixed |
| replace global-feature ranking with distance-based ground-truth ranking | pose approximation (Task 1) · datasets with retrieval GT | % localized (5m, 10deg) | Improved |
| increase k for local SFM map construction (k = 5, 10, 20) | local SFM (Task 2a) · multiple datasets | % localized (0.25m, 2deg): k = 10 good compromise | Improved |
| replace global-feature retrieval with co-observation GT ranking | local SFM (Task 2a) · datasets with retrieval GT | % localized (0.25m, 2deg) | Improved |
| increase k for global SFM (about k = 20) | global SFM (Task 2b) · multiple datasets | % localized (0.25m, 2deg): about k = 20 for top performance | Improved |
| use AP-GeM descriptor instead of NetVLAD/DenseVLAD | local SFM (Task 2a) · Aachen, RobotCar night | % localized (0.25m, 2deg) | Improved |
| use learned descriptors (AP-GeM, DELG-GLDv2) instead of DenseVLAD/NetVLAD for night queries | global SFM (Task 2b) · Aachen night, RobotCar night | % localized (0.25m, 2deg) | Improved |
| use DenseVLAD (SIFT-based) instead of learned descriptors | global SFM (Task 2b) · InLoc DUC2 | % localized (0.25m, 2deg) | Improved |
| use DenseVLAD instead of learned descriptors on night queries | pose approximation (Task 1) · RobotCar night, Aachen night | % localized (5m, 10deg) | Hurt |
Closest papers by meaning
- Long-Term Visual Localization Revisited (TPAMI 2020 · similarity 0.91)
- Yes, we CANN: Constrained Approximate Nearest Neighbors for Local Feature-Based Visual Localization (ICCV 2023 · similarity 0.91)
- A Guide to Structureless Visual Localization (IJCV 2026 · similarity 0.90)
- RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations (CVPR 2025 · similarity 0.90)
- On the Estimation of Image-matching Uncertainty in Visual Place Recognition (CVPR 2024 · similarity 0.88)
- Weatherproofing Retrieval for Localization with Generative AI and Geometric Consistency (ICLR 2024 · similarity 0.88)
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.
