AANet

This repository provides the code for ICRA2023 paper "AANet: Aggregation and Alignment Network with Semi-hard Positive Sample Mining for Hierarchical Place Recognition".

The usage of this repo is similar to the Visual Geo-localization Benchmark. You can refer to it to prepare datasets.

Training

For MSLS

python3 train.py --datasets_folder=/path/to/your/datasets/folder --dataset_name=msls --queries_per_epoch=40000 --trunc_te=8 --freeze_te=1 --negs_num_per_query=2 --candipositive_global=0.3 --candipositive_local=0.3

For Pitts30k

python3 train.py --datasets_folder=/path/to/your/datasets/folder --dataset_name=pitts30k  --trunc_te=8 --freeze_te=1 --negs_num_per_query=2 --horizontal_flip --random_resized_crop=0.3 --candipositive_global=1 --candipositive_local=2 --resume /path/to/your/trained/msls/model/msls.pth

Evaluation

For MSLS

python3 eval.py --datasets_folder=/path/to/your/datasets/folder --dataset_name=msls --trunc_te=8 --freeze_te=1 --resume /path/to/your/trained/msls/model/msls.pth

For Pitts30k

python3 eval.py --datasets_folder=/path/to/your/datasets/folder --dataset_name=pitts30k --trunc_te=8 --freeze_te=1 --resume /path/to/your/trained/pitts30k/model/pitts30k.pth

Trained models

The model for MSLS.

The model for Pitts30k.

Citation

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

@INPROCEEDINGS{aanet,
  author={Lu, Feng and Zhang, Lijun and Dong, Shuting and Chen, Baifan and Yuan, Chun},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={AANet: Aggregation and Alignment Network with Semi-hard Positive Sample Mining for Hierarchical Place Recognition}, 
  year={2023},
  pages={11771-11778},
  doi={10.1109/ICRA48891.2023.10160734}}

Acknowledgements

The structure of this repo is based on Visual Geo-localization Benchmark.