cmSalGAN_PyTorch

Official PyTorch implementation of cmSalGAN: RGB-D Salient Object Detection with Cross-View Generative Adversarial Networks [Project-page] [arXiv]， which is accepted by IEEE Transactions on Multimedia 2020.

Introduction

In this repo, we provide cmSalGAN's code. The discriminator is designed based on SalGAN: Visual Saliency Prediction with Generative Adversarial Networks (CVPR 2017 SUNw). The attention mechanism is designed based on PiCANet: Pixel-wise Contextual Attention Learning for Accurate Saliency Detection (CVPR 2018).

Dependencies

---

• Pytorch 0.4.1+
• python3.5

Training

---

Preparing the training set (Google drive) in Dataset/train/, which contains the subfolders GT, RGB, depth.

Preparing the edge dataset (Google drive) in Dataset/train/.

Then run the command:

python3 train.py

Test

---

Preparing the test set in Dataset/. (Note: The depth images need to be adjusted to 0-255 by normDepth.m )

Downloading the pretrained model (Google drive) to the path models/.

Then run the command:

python3 test.py

Evaluation Results

---

We provide the saliency maps (Google drive) for NJUD, NLPR and STEREO.

We also provide the saliency maps (Google drive) for DES, LFSD, SSD, DUT-RGBD and SIP.

The evaluation-tool we used : https://github.com/wenguanwang/SODsurvey/

Cite

If you find this paper useful for your research, please cite following papers:

@article{jiang2019cmsalgan,
  title={cmSalGAN: RGB-D Salient Object Detection with Cross-View Generative Adversarial Networks},
  author={Jiang, Bo and Zhou, Zitai and Wang, Xiao and Tang, Jin and Bin, Luo},
  journal={IEEE Transactions on Multimedia},
  doi={DOI: 10.1109/TMM.2020.2997184}, 
  year={2020}
}