This repository is initially based on ssd.pytorch, and some modules are from RFBNet and deformable-convolution-pytorch, a huge thank to them.
We currently only support PyTorch-0.4.0 and CUDA 8.0. ./make.sh is required for DeformableConv and COCO tools during installation.
Our models are aviliable at GoogleDrive and BaiduYun
cd script
./train.sh
./train_trn.sh
cd script
./batch_eval.sh
./batch_eval_trn.sh
arXiv paper describes this project in detail.
Xingyu Chen, Junzhi Yu, Shihan Kong, Zhengxing Wu, and Li Wen, "Towards Real-Time Accurate Object Detection in Both Images and Videos Based on Dual Refinement", arXiv:1807.08638, 2018.