CST_DA_detection

The pytorch implementation of paper "Collaborative Training between Region Proposal Localization and Classification for Domain Adaptive Object Detection" in ECCV2020.

Introduction

The project is based on this repository. Please follow the instructions to setup the enviroment. We used Pytorch 0.4.0 and torchvision 0.2.2 for this project.

Data Preparation

• Cityscapes Please refer to this website to download the data. The training set includes 2964 images.
• FoggyCityscapes Please refer to this website to download the data.

Unzip both dataset in ./data, then run boxes_for_cityscapes.py to convert the annotation format.

Training and Evaluation

1. Pretrain with source data.

   CUDA_VISIBLE_DEVICES=${gpu_id} python cityscapes_pretrain.py

2. Domain adaptation with target data.

   CUDA_VISIBLE_DEVICES=${gpu_id} python cityscapes_to_foggycityscapes_da.py

3. Test with target data.

   CUDA_VISIBLE_DEVICES=${gpu_id} python cityscapes_to_foggycityscapes_da_test.py --checksession ${session_id} --checkepoch ${epoch_num} --checkpoint ${point_num}

Pretrained Model

The pretrained model will be released soon.

Citation

Please cite the following reference if you utilize this repository for your project.

@article{zhao2020collaborative,
  title={Collaborative Training between Region Proposal Localization and Classification for Domain Adaptive Object Detection},
  author={Zhao, Ganlong and Li, Guanbin and Xu, Ruijia and Lin, Liang},
  journal={arXiv preprint arXiv:2009.08119},
  year={2020}
}