CrossCL

Official PyTorch implementation of ICME2023 paper “Self-supervised Cross-stage Regional Contrastive Learning for Object Detection”. Paper.

News

• 2023/08/17 We apply CrossCL ResNet50 backbone to our sparse detector ASAG and improve +1.2 AP.
• 2023/07/26 Checkpoints and Logs are released at here, key: icme.
• 2023/04/02 Code released.

Environments

• python 3.7
• pytorch 1.6.0
• cuda 10.2

Pre-training

All the instructions for pre-training a ResNet50-FPN on ImageNet or COCO can be found in ./sh.

Please modify the path to the dataset according to your local path.

For example, to pre-train CrossCL on ImageNet for 200 epochs, run the following command:

bash sh/crosscl_200e_imagenet.sh

Please note that by default, the model is trained on 8 GPUs.

Transferring to Object Detection

To transfer with Detectron2, you should convert the pre-trained model to a standard R50-FPN model by running the following command:

python tools/convert-pretrain-to-detectron.py /path/to/input/checkpoint /path/to/output/checkpoint

Then, you can use the official Detectron2 to train the detection model.

Note, you should first install Detectron2 and then place COCO2017 dataset to detection/datasets/coco (you may use a soft-link).

Finally, to train a Mask R-CNN on COCO for 1x schedule, run the following script:

cd detection
bash finetune_coco_1x.sh

To transfer with mmdetection, you should convert the pre-trained model to a standard R50 model by running the following command:

python tools/convert-pretrain-to-mmdetection.py /path/to/input/checkpoint /path/to/output/checkpoint

Afterwards, you can go ahead to use the official mmdetection to train the detection model.

Note, you should first install mmdetection.

To train a Mask R-CNN on COCO for 1x schedule, run the following training script:

cd mmdetection
sh ./tools/dist_train.sh configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py \ 
    --options model.init_cfg.checkpoint=/path/to/output/checkpoint 8

Note, we do not provide the training code for training with mmdetection.

However, you can use the converted backbone model and follow the guidelines of mmdetection to train the model.

Checkpoints and Logs

The pre-training/converted checkpoints and the pre-training/fine-tunig training logs can be downloaded here, key: icme.

If you have any problem, plz feel free to open an issue or contact me.

Acknowledgement

• This repository is heavily based on MoCo and ReSim.

• If you use this paper/code in your research, please consider citing us:

@inproceedings{yan2023cross,
  title={Self-supervised Cross-stage Regional Contrastive Learning for Object Detection},
  author={Yan, junkai and Yang, Lingxiao and Gao, Yipeng and Zheng, Wei-Shi},
  booktitle={ICME},
  year={2023},
}