Saliency-Aware Texture Smoothing

by Lei Zhu*, Xiaowei Hu*, Chi-Wing Fu, Jing Qin and Pheng-Ann Heng (* joint first authors)

Guided Non-local Block for Saliency Detection

This implementation is written by Xiaowei Hu at the Chinese University of Hong Kong.

---

Citation

@article{zhu2020saliency,   
  author = {Zhu, Lei and Hu, Xiaowei and Fu, Chi-Wing and Qin, Jing and Heng, Pheng-Ann},    
  title = {Saliency-Aware Texture Smoothing},    
  journal={IEEE Transactions on Visualization and Computer Graphics},         
  volume={26},                      
  number={7},              
  pages={2471-2484},      
  year={2020}
}

Dataset

Saliency Dataset for Texture Smoothing (SDTS) can be downloaded from Google Drive.

Installation

1. Please download and compile our CF-Caffe.

2. Clone the GNLB repository, and we'll call the directory that you cloned as GNLB-master.

   git clone https://github.com/xw-hu/GNLB.git

3. Replace CF-Caffe/examples/ by GNLB-master/examples/. Replace CF-Caffe/data/ by GNLB-master/data/.

Test

1. Enter the examples/GNLB/ and run test_saliency.m in Matlab.

2. Apply CRF to do the post-processing for each image.
   The code for CRF can be found in https://github.com/Andrew-Qibin/dss_crf
   *Note that please provide a link to the original code as a footnote or a citation if you plan to use it.

Train

1. Download the pre-trained ResNet-101 caffemodel on ImageNet.
   Put this model in CF-Caffe/models/.

2. Enter the examples/GNLB/GNLB/
   Modify the image path in train_val.prototxt.
   Modify the weight path in train.sh for different training sets (MSRA10K or SDTS) following our paper.

3. Run

   sh train.sh