Research map / Label-efficient and robust learning / Recognition and long-tail
Stochastic Attraction-Repulsion Embedding for Large Scale Image Localization
Liu Liu, Hongdong Li, Yuchao Dai
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 |
|---|---|---|---|
| train NetVLAD with Gaussian SARE (Our-Ind.) instead of standard NetVLAD training | NetVLAD · Sf-0 | Recall@1 (%): 75.58 -> 80.60 | Improved |
| train NetVLAD with Gaussian SARE (Our-Ind.) instead of CRN's best | NetVLAD · 24/7 Tokyo | Recall@1 (%): 75.20 -> 79.68 | Improved |
| train with Gaussian SARE (Our-Ind.) instead of standard NetVLAD training | NetVLAD · Pitts250k-test | Recall@1 (%): 85.95 -> 88.97 | Improved |
| train with Gaussian SARE (Our-Joint) instead of standard NetVLAD training | NetVLAD · TokyoTM-val | Recall@1 (%): 93.85 -> 94.71 | Improved |
| handle multiple negatives jointly instead of independently (ablation) | SARE · Pitts250k-test | Recall@1 (%): 88.43 vs 88.97 | No clear effect |
| handle multiple negatives jointly instead of independently (ablation) | SARE · 24/7 Tokyo | Recall@1 (%): 79.68 vs 80.63 | No clear effect |
| handle multiple negatives jointly instead of independently (ablation) | SARE · Sf-0 | Recall@1 (%): 80.60 -> 77.75 | Hurt |
| use Gaussian kernel instead of Cauchy or Exponential kernel in SARE (ablation) | SARE · Pitts250k-test, TokyoTM-val, 24/7 Tokyo, Sf-0 | Recall | Improved |
| use Gaussian SARE (Our-Ind.) instead of Contrastive loss | contrastive loss · Sf-0 | Recall@1 (%): 74.63 -> 80.60 | Improved |
| use Gaussian SARE (Our-Ind.) instead of Geo-classification loss | Geo-classification loss · 24/7 Tokyo | Recall@1 (%): 71.43 -> 79.68 | Improved |
| use Gaussian SARE (Our-Ind.) instead of N-pair / Angular / Lifted / Ratio loss | metric learning losses · Pitts250k-test, TokyoTM-val, 24/7 Tokyo, Sf-0 | Recall@1 (%): Our-Ind. 88.97 vs Angular 88.60 on Pitts250k-test | No clear effect |
| train with SARE (Our-Ind.) and test on image retrieval without fine-tuning | NetVLAD · Oxford 5k (full) | mAP: 69.09 -> 71.66 | Improved |
Closest papers by meaning
- Yes, we CANN: Constrained Approximate Nearest Neighbors for Local Feature-Based Visual Localization (ICCV 2023 · similarity 0.88)
- Stochastic positional embeddings improve masked image modeling (ICML 2024 · similarity 0.86)
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.86)
- RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations (CVPR 2025 · similarity 0.85)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.85)
- Scaling Image Geo-Localization to Continent Level (NeurIPS 2025 · similarity 0.84)
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.
