Introduction

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.

Environment

The environment required is exactly the same as https://github.com/roytseng-tw/Detectron.pytorch

Datasets

1、Visdrone http://www.aiskyeye.com/upfile/Vision_Meets_Drones_A_Challenge.pdf

2、UAVDT https://sites.google.com/site/daviddo0323/projects/uavdt

Backbone

We choose e2e_mask_rcnn_R-101-FPN_1x.yaml and e2e_mask_rcnn_X-152-32x8d-FPN-IN5k_1.44x.

How to train

bash zjy_train.sh

How to test

bash zjy_test.sh

Result

The results in dataset Visdrone:

The results in dataset UAVDT:

Introduction of related files

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.