Research map / Vision-language and multimodal / Cross-modal learning
CricaVPR: Cross-image Correlation-aware Representation Learning 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 CricaVPR, a VPR global descriptor that correlates images within a batch via a cross-image transformer encoder and adapts frozen DINOv2 with multi-scale convolution adapters.
Problem
Visual place recognition global descriptors are computed from each image in isolation, ignoring cross-image variations (viewpoint, illumination) and leaving them vulnerable to condition changes and perceptual aliasing; full fine-tuning of foundation models also causes forgetting.
Approach
Attention across images in a training batch: the i-th regional features of all images in a batch form a sequence fed to a cross-image encoder, so each image's descriptor absorbs information from others. Paired with a MulConv adapter (1x1/3x3/5x5 parallel convolutions) inserted in parallel to MLP layers of a frozen DINOv2 backbone.
Key results
- CricaVPR (4096-dim) reaches 94.9% R@1 on Pitts30k, 93.0% R@1 on Tokyo24/7 and 90.0% R@1 on MSLS-val, outperforming EigenPlaces and MixVPR.
- On challenging datasets CricaVPR attains 90.7 R@1 on Nordland, 64.7 on AmsterTime and 85.1 on SVOX-Night, improving on the best prior method by 14.5, 15.8 and 20.7 points.
- Training takes 3.5 hours on two RTX 3090 GPUs; with only 0.1 epoch on 10% of data it already beats previous methods except EigenPlaces on Pitts30k and Tokyo24/7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-image correlation awareness to AdaptSPM (ablation) | CricaVPR · Pitts30k | R@1: 90.6 -> 94.8 | Helped |
| add cross-image correlation awareness to AdaptSPM (ablation) | CricaVPR · Tokyo24/7 | R@1: 85.1 -> 93.0 | Helped |
| add cross-image correlation awareness to AdaptSPM (ablation) | CricaVPR · MSLS-val | R@1: 85.5 -> 89.9 | Helped |
| adapt the frozen DINOv2 backbone with MulConvAdapter (ablation) | CricaVPR using GeM features · Pitts30k | R@1: 79.2 -> 87.1 | Helped |
| adapt the frozen DINOv2 backbone with MulConvAdapter (ablation) | CricaVPR using GeM features · Tokyo24/7 | R@1: 65.4 -> 70.2 | Helped |
| adapt the frozen DINOv2 backbone with MulConvAdapter (ablation) | CricaVPR using GeM features · MSLS-val | R@1: 40.8 -> 78.4 | Helped |
| fully fine-tune DINOv2 rather than freeze it (ablation) | CricaVPR · Pitts30k | R@1: 79.2 -> 94.1 | Helped |
| fully fine-tune DINOv2 rather than freeze it (ablation) | CricaVPR · Tokyo24/7 | R@1: 80.0 -> 76.8 | Hurt |
| fully fine-tune DINOv2 rather than freeze it (ablation) | CricaVPR · MSLS-val | R@1: 58.8 -> 86.2 | Helped |
| use MulConvAdapter instead of VanillaAdapter (ablation) | CricaVPR · Pitts30k | R@1: 94.6 -> 94.8 | Helped |
| use MulConvAdapter instead of ConvAdapter (ablation) | CricaVPR · MSLS-val | R@1: 88.0 -> 89.9 | Helped |
| use 3×3 ConvAdapter instead of VanillaAdapter (ablation) | CricaVPR · Pitts30k | R@1: 94.6 -> 93.8 | Hurt |
Surprising findings
- Convpass-style 3x3 convolution adapter degrades VPR performance relative to a vanilla adapter, suggesting improper local priors hurt.
- Full fine-tuning of DINOv2 causes catastrophic forgetting, performing worse than frozen DINOv2 on Tokyo24/7.
Limitations
- Cross-image encoder relies on batch composition, so retrieval uses single-image descriptors only while training depends on batches of correlated images.
- Gains on Tokyo24/7 and MSLS depend on higher-dimensional descriptors; low-dimensional (512-dim) versions degrade notably there.
Experimental setup
Datasets: GSV-Cities
Benchmarks: Pitts30k Tokyo24/7 MSLS-val MSLS-challenge Nordland AmsterTime SVOX
Metrics: Recall@1 Recall@5 Recall@10
Baselines: NetVLAD SFRS CosPlace GCL MixVPR EigenPlaces Patch-NetVLAD TransVPR AnyLoc
Builds on: DINOv2 Convpass GSV-Cities MixVPR multi-similarity loss
Closest papers by meaning
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.88)
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.88)
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · similarity 0.87)
- DistilVPR: Cross-Modal Knowledge Distillation for Visual Place Recognition (AAAI 2024 · similarity 0.87)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.87)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.87)
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.