Super Diffusion for Salient Object Detection(TIP2019)

Peng Jiang, Zhiyi Pan, Changhe Tu, Nuno Vasconcelos, Baoquan Chen, and Jingliang Peng

Introduction

In this work, we firstly present a novel view of the working mechanism of the diffusion process based on mathematical analysis, which reveals that the diffusion process is actually computing the similarity of nodes with respect to the seeds based on diffusion maps. Following this analysis, we propose super diffusion, a novel inclusive learning-based framework for salient object detection, which makes the optimum and robust performance by integrating a large pool of feature spaces, scales and even features originally computed for non-diffusion-based salient object detection.

Performance Statistics of Unsupervised/Supervised Algorithms

For each dataset and protocol, the top three results are highlighted in red, blue and green, respectively. The ↑/↓ sign indicates that the value is positively/negatively related with the performance.

Tabel 1. performance statistics of unsupervised algorithms

Tabel 2. performance statistics of supervised algorithms

Usage

1. Preparation:

   • Clone the repository. This code is tested on the MATLAB R2017a.

   • Run the make.m in EntropyRateSuperpixel and put the mex_ers in root folder.

   • addpath('./others/')

2. Dataset

   Download the MSRA10K and change the imgRoot to your MSRA10K path in SuperDiffusion.m.

3. Evaluation

   • Change the saldir to your save folder in SuperDiffusion.m.

   • Put the ./weight.mat in your save folder.

   • run SuperDiffusion.m

4. Training

   • Change the saldir to your save folder in SuperDiffsion.m.

   • If you want to use saliency features computed by other methods as input, put features in ./otherfeature and change the args.otherfeature in SuperDiffusion.m.

   • run SuperDiffusion.m

Citation

If SuperDiffusion is useful for your research, please consider citing:

@article{DBLP:journals/tip/JiangPTVCP20,
  author    = {Peng Jiang and
               Zhiyi Pan and
               Changhe Tu and
               Nuno Vasconcelos and
               Baoquan Chen and
               Jingliang Peng},
  title     = {Super Diffusion for Salient Object Detection},
  journal   = {{IEEE} Trans. Image Processing}
}