This repo contains the official implementation of the ACM MM 2024 paper
MFRGN: Multi-scale Feature Representation Generalization Network for Ground-to-Aerial Geo-localization
paper(MM'24)
This code is based on the Sample4Geo framework.
Details of the datasets, training and inference can be found in Sample4Geo.
- 12 Dec 2024 We now provide supplementary results on University-1652.
| Methods | Drone2Sat R@1 / AP |
Sat2Drone R@1 / AP |
|---|---|---|
| Sample4Geo | 92.65 / 93.81 | 95.65 / 91.39 |
| Ours | 94.33 / 95.24 | 96.15 / 93.94 |
To accelerate training/test time, you can run data_preparation.py， which implements image transformation (from '.jpg'/'.png' to '.pt') and cropping (similar to SAFA).
When you process images from '.jpg'/'.png' to '.pt', you should set ext='pt' in sample4geo/dataset/*.py
Also, if you are experiencing network errors about the backbone, you may need to download the backbone weights offline and put them into the pretrained folder.
We provide our pretrained results: MFRGN-pretained.zip [BaiduYun, Password: 1234], which contains pretrained weight files or files necessary to train certain network configurations (e.g. distance_dict, convnext backbone weights).