- CVUSA：https://github.com/viibridges/crossnet
- CVACT：https://github.com/Liumouliu/OriCNN
Our method follow SAFA and DSM to get polar transform aerial images. Please download data_prearation.py for pre-processing.
Pretrained model Download(Google's Official Checkpoint)
- R50+ViT-B_16 pretrained model
Our trained models for CVUSA and CVACT will be available soon
Train
python train.py --name CVUSA --dataset CVUSA --pretrained_dir YOUR_PRETRAINED_MODEL --output_dir MODEL_WHERE_WILL_BE_SAVE --dataset_dir YOUR_DATASET_PATH --learning_rate 1e-4 --weight_decay 0.03
Test
python test.py --name CVUSA --dataset CVUSA --output_dir MODEL_WHERE_WILL_BE_SAVE --dataset_dir YOUR_DATASET_PATH
If you want to use auto mixed precision, you should install APEX and add --fp16 in startup code.
| dataset | top-1 | top-5 | top-10 | top-1% |
|---|---|---|---|---|
| CVUSA | 94.05% | 98.27% | 98.99% | 99.67% |
| CVACT_val | 84.89% | 94.59% | 95.96% | 98.37% |