Research map / Label-efficient and robust learning / Recognition and long-tail
MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering
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 MutualVPR, which co-evolves a DINOv2-based descriptor encoder and K-means clustering of view-direction categories within UTM grid cells, removing reliance on orientation labels, and reports SOTA-competitive VPR recall on SF-XL-trained benchmarks.
Problem
Classification-based visual place recognition suffers from supervision inconsistency: fixed view-direction labels (CosPlace) or crop-based grouping (EigenPlaces) assign visually dissimilar views to the same class or similar views to different classes, especially under occlusion.
Approach
Alternate between descriptor learning and K-means reclustering of intra-grid view categories, with a DINOv2 encoder (adapter-tuned) and LMCL classifier, so place labels are refined by learned visual similarity instead of fixed orientation labels.
Key results
- On SF-XL-Occlusion, MutualVPR reaches 47.4% R@1 versus 36.8% for EigenPlaces and 32.9% for CosPlace.
- On SF-XL-testv1, MutualVPR gets 80.8% R@1 versus 64.8% for CosPlace, but EigenPlaces scores higher at 83.8%.
- On Tokyo24/7, MutualVPR reaches 92.4% R@1 among classification methods, but the paper's Tokyo result for the DINOv2 ablation is 92.1% R@1 with 0+30 degree cropping.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive clustering instead of fixed view-direction labels | VPR classification training · SF-XL-Occlusion | R@1: 32.9 -> 47.4 | Helped |
| use adaptive clustering instead of fixed view-direction labels | VPR classification training · SF-XL-Occlusion | R@1: 36.8 -> 47.4 | Helped |
| add intra-grid classification rather than no classification (ablation) | MutualVPR · SF-XL-Occlusion | R@1: 17.1 -> 47.4 | Helped |
| add multi-angle cropping (0°+30°) (ablation) | CosPlace with ResNet50 · SF-XL-testv1 | R@1: 64.8 -> 81.1 | Helped |
| add multi-angle cropping (0°+30°) (ablation) | CosPlace with ResNet50 · Tokyo24/7 | R@1: 76.5 -> 85.1 | Helped |
| add multi-angle cropping (0°+30°) (ablation) | MutualVPR with ResNet50 · Tokyo24/7 | R@1: 82.9 -> 85.4 | Helped |
| add multi-angle cropping (0°+30°) (ablation) | MutualVPR with DINOv2 · SF-XL-testv1 | R@1: 77.0 -> 80.8 | Helped |
| increase cluster count from K=1 to K=3 (ablation) | MutualVPR with ResNet50 · Tokyo24/7 | R@1: 68.3 -> 81.0 | Helped |
| increase cluster count from K=3 to K=6 (ablation) | MutualVPR with ResNet50 · Tokyo24/7 | R@1: 81.0 -> 82.9 | Helped |
| increase cluster count from K=3 to K=6 (ablation) | MutualVPR with DINOv2 · Tokyo24/7 | R@1: 91.1 -> 86.0 | Hurt |
| increase cluster count from K=3 to K=6 (ablation) | MutualVPR with DINOv2 · SF-XL-testv1 | R@1: 77.0 -> 70.9 | Hurt |
Surprising findings
- Without any intra-grid classification (K=1), DINOv2 performance drops sharply (SF-XL-testv1 R@1 61.1%), showing the clustering is essential.
- For DINOv2, K=3 outperforms K=6 by roughly 10 points, so the optimal cluster number depends on the backbone.
Limitations
- Fixed cluster number K must be hand-chosen and its optimum depends on backbone and dataset.
- Does not consistently beat EigenPlaces on SF-XL-testv1 and Pitts30k/Pitts250k; not always top performer.
- Relies on hand-set grid size M and cropping angles (0 and 30 degrees).
Experimental setup
Datasets: SF-XL GSV-Cities Pitts30k-test Pitts250k-test MSLS-val Tokyo 24/7 SF-XL-test v1 SF-XL-occlusion
Benchmarks: MSLS-val Pitts30k-test Pitts250k-test Tokyo 24/7 SF-XL-test v1 SF-XL-Occlusion
Metrics: Recall@1 Recall@5 Recall@10 Recall@20
Baselines: NetVLAD GeM AnyLoc ConvAP MixVPR CricaVPR BoQ SALAD SALAD+CM EigenPlaces CosPlace
Builds on: CosPlace EigenPlaces DINOv2 LMCL MulConv adapter
Closest papers by meaning
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.90)
- Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · similarity 0.89)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.89)
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.88)
- CricaVPR: Cross-image Correlation-aware Representation Learning for Visual Place Recognition (CVPR 2024 · similarity 0.88)
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · 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.