This is the codes of ICCV workshop paper "How to Fully Exploit The Abilities of Aerial Image Detectors". Our model is based on detectron.pytorch.
The environment required is exactly the same as https://github.com/roytseng-tw/Detectron.pytorch
1、Visdrone http://www.aiskyeye.com/upfile/Vision_Meets_Drones_A_Challenge.pdf
2、UAVDT https://sites.google.com/site/daviddo0323/projects/uavdt
We choose e2e_mask_rcnn_R-101-FPN_1x.yaml and e2e_mask_rcnn_X-152-32x8d-FPN-IN5k_1.44x.
bash zjy_train.sh
bash zjy_test.sh
The results in dataset Visdrone:
The results in dataset UAVDT:
1、prop.json This is the predicted difficult regions on testset.
2、lib/roi_data/fast_rcnn.py The IoU-balanced sampling is add to here.
3、lib/utils/net.py The balanced L1 losss is add to here.
4.SSD Difficult region network and the related tool files.