Home

Awesome

DetCo: Unsupervised Contrastive Learning for Object Detection

arxiv link

News

Highlights

Pipeline

image-20190807160835333

Performances

Graph


Graph


Graph


Graph


Graph

Install

Same as OpenSelfSup.

Codes

Pretext Task Pretrain

Coming Soon.

Transfer to Downstream tasks

We provide training scripts on COCO, because the performance of COCO is more stable than VOC and Cityscapes. See results in Table 3-5 and Table 13.

We provide Mask RCNN-C4, Mask RCNN-FPN and RetinaNet with 12k, 90k and 180k iterations.

First, you need to download model(.pkl) to benchmarks/detection/pths, and convert pretrain model to detectron2_version. See this script.

Second, start training and testing.

sh tools_local/dist_test_coco.sh $PTH $WORK_DIR

For example:

sh tools_local/dist_test_coco.sh benchmarks/detection/pths/detco_200ep_AA.pkl benchmarks/detection/work_dirs/detco_AA

Download Models

DetCo-200ep: [Google Drive], [Baidu Drive] Fetch Code: okfp

DetCo-200ep-AA: [Google Drive], [Baidu Drive] Fetch Code: fg7h

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

@misc{xie2021detco,
      title={DetCo: Unsupervised Contrastive Learning for Object Detection}, 
      author={Enze Xie and Jian Ding and Wenhai Wang and Xiaohang Zhan and Hang Xu and Zhenguo Li and Ping Luo},
      year={2021},
      eprint={2102.04803},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledges

We would like to thank Huawei AI Theory Group to support 200+ V100 GPUs for this research project without which this work would not be possible.

License

For academic use, this project is licensed under the 2-clause BSD License - see the LICENSE file for details. For commercial use, please contact the authors.