This repo. is an official implementation of the DPANet , which has been accepted in the journal IEEE Transactions on Image Processing, 2020.

The main pipeline is shown as the following,

And some visualization results are listed

>= Pytorch 1.0.0
OpenCV-Python
[optional] matlab

• download the official pretrained model of ResNet-50/ResNet-34/ResNet-18 implemented in Pytorch if you want to train the network again.
• download or put the RGB-D saliency benchmark datasets (Google drive) in the folder of data for training or test.
• [optional] generate the pseudo label (provided for NJUD and NLPR) using the scripts gen.m and cal_score.py.

python3 train.py --tag res50 --save_path res50_res

make sure that the GPU memory is enough (the original training is conducted on 8 NVIDIA RTX2080Ti cards with the batch size of 32).

 python3 test.py --tag res50 --gpu 0 --model res50_res/model-30

We provide the trained model file (Google drive), and run this command to check its integrity:

md5sum model-res50-epoch30.pt 

you will obtain the code b666d297e0237035f6e48f80711ca927. Please use the matlab code to evaluate the MAE, F-measure, or other metrics rather than using the accuracy defined in the test.py.

The saliency maps are also available (Google drive).

We provide the evaluation code in the folder "eval_code" for fair comparisons. You may need to revise the algorithms , prepath, and maskpath defined in the main.m. The saliency maps of the competitors (official maps or obtained by running the official code) are provided (Google drive).

Please cite the DPANet in your publications if it helps your research:

@article{chen2020depth,
  title={{DPANet}:Depth Potentiality-Aware Gated Attention Network for {RGB-D} Salient Object Detection},
  author={Chen, Zuyao and Cong, Runmin and Xu, Qianqian and Huang, Qingming},
  journal={IEEE Transactions on Image Processing},
  year={2020}
}