Awesome

TransGeo2022

This repository provides the code for "TransGeo: Transformer Is All You Need for Cross-view Image Geo-localization".

@inproceedings{zhu2022transgeo,
  title={TransGeo: Transformer Is All You Need for Cross-view Image Geo-localization},
  author={Zhu, Sijie and Shah, Mubarak and Chen, Chen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1162--1171},
  year={2022}
}

Dataset

Please prepare VIGOR, CVUSA or CVACT. You may need to modify specific path in dataloader.

Requirement

- Python >= 3.6, numpy, matplotlib, pillow, ptflops, timm
- PyTorch >= 1.8.1, torchvision >= 0.11.1

Training and Evaluation

Simply run the scripts like:

sh run_CVUSA.sh

You may need to specify the GPUs for training in "train.py". Remove the second line if you want to train the simple stage-1 model. Change the "--dataset" to train on other datasets. The code follows the multiprocessing distributed training style from PyTorch and Moco, but it only uses one GPU by default for training. You may need to tune the learning rate for multi-GPU training, e.g. linear scaling rule. There may be memory issue if training with too many GPUs with our sampler.

We follow timm, ViT and Deit for pytorch implementation of vision transformer. We use the pytorch implementation of ASAM.

Pretrained Models

You can find the pretrained models here. Simply add "-e" for train.py if you want to evaluate with pretrained weights. Please refer to VIGOR for meter-level evaluation code. Validation features for VIGOR are also available for quick meter-level evaluation from npy. Let me know if there is anything wrong with the link.

Reference

- http://mvrl.cs.uky.edu/datasets/cvusa/
- https://github.com/Jeff-Zilence/VIGOR
- https://github.com/Liumouliu/OriCNN
- https://github.com/facebookresearch/deit
- https://github.com/facebookresearch/moco
- https://github.com/davda54/sam
- https://github.com/david-husx/crossview_localisation.git