Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition

This is the official repository for the AAAI 2025 paper "FoL" available at AAAI Paper Page. In addition, our paper and its extensive supplementary materials can be found on arXiv.

Summary

We introduce Focus on Local (FoL), a two-stage Visual Place Recognition (VPR) approach that enhances image retrieval and re-ranking by identifying and leveraging reliable discriminative local regions. Our method introduces three key contributions:

• Reliable Discriminative Region Modeling: We propose two novel loss functions—Extraction-Aggregation Spatial Alignment Loss (SAL) and Foreground-Background Contrast Enhancement Loss (CEL)—to explicitly learn discriminative local regions.
• Weakly-Supervised Local Feature Learning: We leverage pseudo-correspondences from aggregated global features to improve local matching supervision.
• Efficient Re-ranking with Discriminative Region Guidance: We use the learned discriminative regions to guide local feature matching, improving accuracy and efficiency.

You may refer to our anonymous conference version of the paper: anonymous conference version

Setup

Tested on Pytorch 2.0.0 with CUDA 11.7. To set up the environment:

conda env create -f environment.yml

Install the Hugging Face Hub client (if you want to pull weights directly):

pip install huggingface_hub

Download Pretrained Weights

You can download our pretrained FoL model either via Google Drive or directly from Hugging Face:

• Google Drive
  here

• Hugging Face Hub

  from huggingface_hub import hf_hub_download
  
  # this will download FoL.pth into your cache folder
  FoLpath = hf_hub_download(
      repo_id="shunpeng/FoL",
      filename="FoL.pth"
  )
  print("Downloaded weights to:", FoLpath)

---

Evaluation

Assuming you have your datasets under /datasets/ and your weights in /weights/FoL.pth:

python eval.py \
  --eval_datasets_folder=/datasets/ \
  --dataset_names pitts30k amstertime \
  --resume=/weights/FoL.pth

Train

python train.py --eval_datasets_folder=.../datasets/ --eval_dataset_name pitts30k --epochs_num=8 --train_batch_size=60 --lr=6e-5 --optim=adamw --resize 322 322 --save_dir train_log/

Performance

This table lists Recall@1, R@5 and R@10 for our FoL model across common VPR datasets, showing results for both FoL-global and FoL-reranking. To facilitate immediate visual comparison, the best R@1 value within each dataset row is typeset in bold.

ID	Dataset	FoL-global			FoL-reranking
		R@1	R@5	R@10	R@1	R@5	R@10
1	Pitts250k-test	96.5	99.1	99.5	97.0	99.2	99.5
2	MSLS-val	93.1	96.9	97.4	93.5	96.9	97.6
3	MSLS-challenge	78.7	90.8	93.0	80.0	90.9	93.0
4	Tokyo24/7	96.2	98.7	98.7	98.4	99.1	99.4
5	Pitts30k	93.9	97.2	98.1	94.5	97.4	98.2
6	Sped	92.1	96.5	98.0	91.8	96.5	97.4
7	Amstertime	64.6	84.3	88.2	70.1	89.0	91.8
8	Eynsham	91.7	95.3	96.2	92.4	95.8	96.6
9	Nordland*	78.3	90.8	94.0	85.8	94.9	96.8
10	Nordland**	87.8	94.5	96.4	92.6	96.9	98.0
11	SF-XL Night	53.4	65.9	71.7	60.5	72.8	75.8
12	SF-XL Occlusion	51.3	65.8	73.7	61.8	77.6	77.6
13	SVOX	98.4	99.4	99.6	98.9	99.6	99.7
14	SVOX Sun	98.1	99.4	99.5	98.8	99.8	99.9
15	SVOX Night	98.3	99.6	99.6	98.8	99.8	99.9
16	SVOX Snow	99.1	99.7	99.8	99.3	99.8	99.9
17	SVOX Overcast	97.9	99.2	99.3	98.3	99.3	99.7
18	SVOX Rain	96.5	99.6	99.7	98.2	99.9	99.9
19	St. Lucia	99.9	100.0	100.0	99.9	100.0	100.0

To-do

• Public release of evaluation code and pretrained FoL model
• Public release of the training code
• More detailed documentation (coming soon)

Acknowledgements

This code is based on the excellent work of:

• DINOv2
• AnyLoc
• SelaVPR
• CricaVPR
• SALAD

Citation

If you find this repo useful for your research, please cite the paper

@inproceedings{wang2025focus,
  title={Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition},
  author={Wang, Changwei and Chen, Shunpeng and Song, Yukun and Xu, Rongtao and Zhang, Zherui and Zhang, Jiguang and Yang, Haoran and Zhang, Yu and Fu, Kexue and Du, Shide and others},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={7},
  pages={7536--7544},
  year={2025}
}