TriSA: A Three Stage Approach for UAV-Satellite Cross-View Geo-Localization Based on Self-Supervised Feature Enhancement
This repository is the code for our paper TriSSA: A Three Stage Strong Approach for UAV-Satellite Cross-View Geo-Localization Based on Self-Supervised Feature Enhancement, Thank you for your kindly attention.
- Download the University-1652 dataset
- Prepare Data Folder
├── University-1652/
│ ├── readme.txt
│ ├── train/
│ ├── drone/ /* drone-view training images
│ ├── 0001
| ├── 0002
| ...
│ ├── street/ /* street-view training images
│ ├── satellite/ /* satellite-view training images
│ ├── google/ /* noisy street-view training images (collected from Google Image)
│ ├── test/
│ ├── query_drone/
│ ├── gallery_drone/
│ ├── query_street/
│ ├── gallery_street/
│ ├── query_satellite/
│ ├── gallery_satellite/
│ ├── 4K_drone/
You can download the trained embedding files (.mat)from the following link and put them in the "evaluaiton/weights/" folder
You can download the trained embedding files (.mat) from the following link and put them in the "evaluaiton/weights/" folder
We prepared the following .py files for easy evaluation:
├── evaluation/
│ ├── evaluate_cpu_no_rerank.py /* test with cpu and no rerank
│ ├── evaluate_cpu_rerank.py /* test with cpu and rerank
│ ├── evaluate_gpu_rerank.py /* test with gpu and rerank
If you are using the gpu-based re-ranking, make sure to compile the file by:
cd evaluation/
sh make.sh
We provide scripts to complete TriSSA training and testing
- Change the data_dir and test_dir paths and then run:
python train.py --gpu_ids 0 --name traied_model_name --train_all --batchsize 32 --data_dir your_data_pathpython test.py --gpu_ids 0 --name traied_model_name --test_dir your_data_path --batchsize 32 --which_epoch 120
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent