CoEGNet

Code for "Re-thinking Co-Salient Object Detection" IEEE TPAMI2021.

Pipeline

Stage 1

1.1 Environments (Caffe && Matlab)

The caffe package is borrowed from https://github.com/BVLC/caffe

1.2 Pre-trained models in Caffe:

• VGG16 model on ImageNet: models/deploy_vgg16CAM.prototxt weights:[http://cnnlocalization.csail.mit.edu/demoCAM/models/vgg16CAM_train_iter_90000.caffemodel]

1.3 test

• simple run ./stage1/demo.m, then we can obtain the initial cosal activations

Stage 2

2.1 Prerequisite

• Python 3.7, PyTorch 1.1.0

2.2 test

• python ./stage2/run_sample.py (Thanks for Ahn et al.'s implement, our postprocessing is inspired by IRNET)

The results of CoEGNET

Baidu Cloud: https://pan.baidu.com/s/19hIlViLbby-a7vQw17ZTVw Fetchcode: f4p3

Google Cloud: https://drive.google.com/file/d/1AK9UNR5mLHQakOTkRawcKSDg8YwIu-xJ/view?usp=sharing

Tips

The sub-net of Co-EGNet is our previous version EGNet. Please refer to the training details in: https://github.com/JXingZhao/EGNet

CoSOD3K dataset: https://github.com/DengPingFan/CoSOD3K

Citation

If you use this code, please cite our paper:

@article{deng2021re,
  title={Re-thinking co-salient object detection},
  author={Deng-Ping, Fan and Tengpeng, Li and Zheng, Lin and Ge-Peng, Ji and Dingwen, Zhang and Ming-Ming, Cheng and Huazhu, Fu and Jianbing, Shen},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  doi={10.1109/TPAMI.2021.3060412}, 
  year={2021}
}