Learning Domain Adaptive Object Detection with Probabilistic Teacher

This repo is the official implementation of ICML2022 paper "Learning Domain Adaptive Object Detection with Probabilistic Teacher" by Meilin Chen, Weijie Chen, Shicai Yang, et al. If you have any problem about this work, please feel free to contact Meilin Chen (merlinis-at-zju.edu.cn) or Weijie Chen (chenweijie5-at-hikvision.com).

Installation

Prerequisites

pip install -r requirements.txt

Install Detectron2

Follow the INSTALL.md to install Detectron2. We use version: detectron2==0.5

Usage

Data Preparation

Plz refer to prepare_data.md for datasets preparation.

Pretrained Model

We used VGG16 pre-trained on ImageNet for all experiments. You can download it to /path/to/project:

• VGG16: Dropbox, VT Server

Training

Plz refer to get_started.md for detailed commands.

Main Results

This code has been further improved, achiving more superior adaptation performance than the results presented in the paper (about +1~2 mAP gains across the tasks, see exps logs for details).

Adaptation Tasks	Methods	Model Weights	mAP50	Log
CitysScape2FoggyCityscape	PT (ours)	Google Drive	31 ⇒ 47.1 (+16.1)	Google Drive
CitysScape2BDD100k	PT (ours)	Google Drive	26.9 ⇒ 34.9 (+8.0)	Google Drive
KITTI2CitysScape	PT (ours)	Google Drive	46.4 ⇒ 60.2 (+13.8)	Google Drive
Sim10k2CitysScape	PT (ours)	Google Drive	44.5 ⇒ 55.1 (+10.6)	Google Drive

Citation

If you use Probabilistic Teacher in your research or wish to refer to the results published in the paper, please consider citing our paper:

@inproceedings{chen2022learning,
  title={Learning Domain Adaptive Object Detection with Probabilistic Teacher},
  author={Chen, Meilin and Chen, Weijie and Yang, Shicai and Song, Jie and Wang, Xinchao and Zhang, Lei and Yan, Yunfeng and Qi, Donglian and Zhuang, Yueting and Xie, Di and others},
  booktitle={International Conference on Machine Learning},
  pages={3040--3055},
  year={2022},
  organization={PMLR}
}

License

This project is released under the Apache 2.0 license. Other codes from open source repository follows the original distributive licenses.

Acknowledgement

This project is built upon Detectron2 and Unbiased Teacher, and we'd like to appreciate for their excellent works.