Research map / Label-efficient and robust learning / Recognition and long-tail
SelaVPR++: Towards Seamless Adaptation of Foundation Models for Efficient Place Recognition
Feng Lu, Tong Jin, Xiangyuan Lan, Lijun Zhang, Yunpeng Liu, Yaowei Wang, Chun Yuan
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 MultiConv adapters in the memory-efficient parallel network instead of vanilla adapters (ablation) | MemEfficient adaptation (DINOv2-B) · Nordland | R@1: 75.0 -> 81.3 | Improved |
| use MultiConv adapters instead of token-channel alternating adapters (ablation) | MemEfficient adaptation (224px) · Nordland | R@1: 73.1 -> 75.2 | Improved |
| use memory-efficient MultiConv adaptation instead of full fine-tuning (ablation) | DINOv2-B backbone · Nordland | R@1: 73.7 -> 81.3 | Improved |
| use memory-efficient MultiConv adaptation instead of frozen backbone (ablation) | DINOv2-B backbone · Nordland | R@1: 53.2 -> 81.3 | Improved |
| use memory-efficient MultiConv adaptation instead of partial tuning (last 4 blocks) (ablation) | DINOv2-B backbone · Nordland | R@1: 77.4 -> 81.3 | Improved |
| use memory-efficient MultiConv adaptation instead of SelaVPR-global adapters (ablation) | DINOv2-B backbone · Nordland | R@1: 73.3 -> 81.3 | Improved |
| use memory-efficient MultiConv adaptation instead of full fine-tuning (ablation) | DINOv2-L backbone · Tokyo24/7 | R@1: 88.9 -> 96.2 | Improved |
| use memory-efficient MultiConv adaptation instead of partial tuning (ablation) | DINOv2-L backbone · Nordland | R@1: 78.8 -> 81.4 | Improved |
| freezing backbone instead of MultiConv adaptation (ablation) | DINOv2-L backbone · Tokyo24/7 | R@5: 98.4 -> 98.4 | No clear effect |
| replace SelaVPR-global adaptation with MemEfficient-MCA to cut training memory (ablation) | DINOv2-L backbone · GPU memory | ratio of memory: 14.6% of SelaVPR-global, 11.5% of Full-tuning | Improved |
| use MultiConv adaptation instead of partial tuning (DINOv2-B) (ablation) | DINOv2-B backbone · GSV-Cities training | min/epoch: 6.85 vs 6.95 | No clear effect |
| apply only 4 adapters instead of 16 for DINOv2-L (ablation) | MemEfficient-MCA · GSV-Cities training | min/epoch: 11.97 -> 9.60 | Improved |
Closest papers by meaning
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · similarity 0.92)
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.91)
- CricaVPR: Cross-image Correlation-aware Representation Learning for Visual Place Recognition (CVPR 2024 · similarity 0.88)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.88)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.88)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 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.
