RAL2025: DSFormer: A Dual-Scale Cross-Learning Transformer for Visual Place Recognition

Introduction

The official repository of "DSFormer: A Dual-Scale Cross-Learning Transformer for Visual Place Recognition".

• Arxiv Preprint
• Official Paper

Environment Setup

• Create conda enviromnet

conda create -n dsformer python=3.10

conda activate dsformer

pip install torch==2.0.1 torchvision==0.15.2 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

• Install rpe-index==1.2.0 (rpe-index) and depend on cuda11.8

cd DSFormer/models/rpe_ops

python setup.py install --user

Testing

• Datasets

  MSLS

  Pittsburgh

  Tokyo24/7

  SFXL

• Trained Models

  We provide pretrained model weights based on both ResNet50 and DINOv2 (ViT-B) backbones, with a global descriptor dimensionality of 512.

  DSFormer (ResNet50): Google Drive

  DSFormer (DINOv2 (ViT-B)): Google Drive

• Run Testing

  python test.py --backbone ResNet50 --dataset_folder /path/to/yourpath --dataset_name msls --split val
  

Block Clustering

If you would like to use the the processed SFXL dataset based on Block Clustering for training your model, we provide two options:

• 1. Directly use datasets/hdbscanplaces_datasets.py, and train_dataset_folder is the path of raw panoramic image dataset (~930G).

  from datasets.hdbscanplaces_dataset import HDBScanPlaces
  groups_num = 5
  groups = [HDBScanPlaces(train_dataset_folder='/path/your/raw_SFXL', 
                          focal_dist=args.focal_dist,
      					current_group=n // 2, 
                          min_images_per_class=10,
      					angle=[0, 90][n % 2])
      					for n in range(groups_num * 2)]
  

• 2. A subset of approximately 4.2 million images(~188G) is extracted from the raw SFXL panoramic dataset using the Block Clustering.

  python extract_images.py --dataset_folder /path/your/raw_SFXL --extracted_folder /path/your/subset
  

  Use:

  from datasets.hdbscanplaces2_dataset import HDBScanPlaces
  groups_num = 5
  groups = [HDBScanPlaces(train_dataset_folder='/path/your/subset', 
      					current_group=n // 2, 
      					angle=[0, 90][n % 2])
      					for n in range(groups_num * 2)]
  

Acknowledgements

EigenPlaces implementation in PyTorch

CosPlace implementation in PyTorch

R2Former implementation in PyTorch

BoQ implementation in PyTorch

IRPE implementation in PyTorch

Cite

If you use this work, please cite:

@ARTICLE{jiang2025dsformer,
  author={Jiang, Haiyang and Piao, Songhao and Gao, Chao and Yu, Lei and Chen, Liguo},
  journal={IEEE Robotics and Automation Letters}, 
  title={DSFormer: A Dual-Scale Cross-Learning Transformer for Visual Place Recognition}, 
  year={2025},
  volume={10},
  number={10},
  pages={10799-10806},
  doi={10.1109/LRA.2025.3604761}}