Research map / Detection and segmentation / CNN features
Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FoL, a two-stage visual place recognition method that models reliable discriminative regions via two losses (SAL, CEL), trains local features with pseudo-correspondences from aggregation clusters, and re-ranks only within discriminative regions.
Problem
Visual place recognition suffers from perceptual aliasing and background distraction; existing two-stage methods do not explicitly model discriminative regions, lack local correspondence labels, and have heavy re-ranking cost.
Approach
Align transformer self-attention with the SALAD Sinkhorn aggregation mask (SAL), enforce foreground-background prototype contrast (CEL), generate pseudo-correspondences from shared cluster assignments for weakly supervised local feature learning, and restrict nearest-neighbor local matching to the top 40% discriminative tokens.
Key results
- FoL-re-ranking reaches 80.0 R@1 on MSLS-challenge vs. 75.0 for SALAD and 73.5 for SelaVPR.
- FoL-global improves R@1 on Tokyo24/7 to 96.2 and Pitts250k-test to 96.5, surpassing one-stage baselines.
- Total single-query re-ranking time on Pitts250k-test is 0.057 s, versus 0.094 s for SelaVPR and 0.356 s for R2Former.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add simple re-ranking (ablation) | Baseline SALAD · MSLS-challenge | R@1: 74.8 -> 75.7 | Helped |
| Add Extraction-Aggregation Spatial Alignment Loss (SAL) (ablation) | FoL baseline with simple re-ranking · MSLS-challenge | R@1: +1.5% | Helped |
| Add Foreground-Background Contrast Enhancement Loss (CEL) (ablation) | FoL with SAL · MSLS-challenge | R@1: +0.9% | Helped |
| Add pseudo-correspondence loss (PC) (ablation) | FoL with SAL and CEL · MSLS-challenge | R@1: +0.6% | Helped |
| Add discriminative region guidance (DRG) for re-ranking (ablation) | FoL with PC loss · MSLS-challenge | R@1: 78.7 -> 80.0 | Helped |
| Use the top 40% of region-mask values for DRG selection (ablation) | FoL re-ranking · VPR benchmarks | Recall@k | Helped |
| Use FoL-global instead of the strongest reported baseline | Pitts250k-test · Pitts250k-test | R@1: 95.7 -> 96.5 | Helped |
| Add FoL re-ranking to global retrieval | FoL-global · Pitts250k-test | R@1: 96.5 -> 97.0 | Helped |
| Use FoL-global instead of SALAD | MSLS-val · MSLS-val | R@1: 92.2 -> 93.1 | Helped |
| Use FoL re-ranking instead of FoL-global | FoL-global · MSLS-challenge | R@1: 78.7 -> 80.0 | Helped |
| Use FoL re-ranking instead of FoL-global | FoL-global · SVOX Rain | R@1: 96.5 -> 98.2 | Helped |
| Use FoL re-ranking instead of FoL-global | FoL-global · St. Lucia | R@1: 99.9 -> 99.9 | No clear effect |
Limitations
- Relies on a fixed top-40% mask threshold k chosen on benchmarks
- Pseudo-correspondence labels are noisy and only weakly supervised
- Only the last four DINOv2 layers are fine-tuned, limiting adaptation capacity
Experimental setup
Datasets: GSV-Cities
Benchmarks: Pitts250k-test MSLS-val MSLS-challenge Tokyo24/7 Nordland AmsterTime SF-XL SVOX SPED Pitts30k Eynsham St. Lucia
Metrics: Recall@1 Recall@5 Recall@10 runtime (s)
Baselines: NetVLAD SFRS Patch-NetVLAD CosPlace MixVPR R2Former EigenPlaces SelaVPR CricaVPR SALAD BoQ TransVPR
Builds on: SALAD SelaVPR DINOv2 multi-similarity loss
Closest papers by meaning
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.91)
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · similarity 0.91)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.90)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.90)
- Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · similarity 0.90)
- On the Estimation of Image-matching Uncertainty in Visual Place Recognition (CVPR 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.