Home

Awesome

Progressive Feature Polishing Network for Salient Object Detection

Contents

Introduction

This repository is the official implementation of PFPN for salient object detection, which is proposed in Progressive Feature Polishing Network for Salient Object Detection.

pfpn

Installation

The code was tested with CentOS7 with Anaconda environment:

Quick demo

Prepare data

Training PFPN

Testing PFPN

After modification of shell/test.sh according to your machine, PFPN testing can be run by sh shell/test.sh.

Evaluation

We implement four metrics: MAE(Mean Absolute Error), F-Measure, S-Measure[2] and E-Measure[1].

Quick evaluation

After preparing data, to make a quick evaluation for PFPN on 5 dataset: ECSSD, HKU-IS, DUTS, DUT-OMRON and PASCAL-S, you can download predicted saliency maps of PFPN(Google Drive or Baidu Netdisk). And then create folder res_vis/PFPN in root directory of this project and extract those results to this folder. The path structure looks like this:

res_vis
├──── PFPN
│    ├──── ECSSD
│    ├──── HKU-IS
│    ├──── DUTS
│    ├──── DUT-OMRON
│    └──── PASCAL-S

Finally, run shells/evaluation.sh to start evaluating:

API

You can use Evaluation in utils/evaluation.py to calculate these four metrics, for example:

    saliency_evaluation = SaliencyEvaluation()
    saliency_evaluation.clear()
    # pred is predicted saliency map, gt is groundtruth
    saliency_evaluation.add_one(pred, gt)
    # get_evaluation would return mean results of all metrics
    MAE, Precision, Recall, F_m, S_m, E_m = saliency_evaluation.get_evaluation()

Experiments

Quantitative Comparison

metric

Qualitative Comparison

comp

Citation

@article{pfpn,
  title = {Progressive Feature Polishing Network for Salient Object Detection},
  author = {Bo, Wang and Quan, Chen and Min, Zhou and Zhiqiang, Zhang and Xiaogang, Jin and Kun, Gai},
  booktitle = {arXiv:1911.05942v1},
  year = {2019}
}

Reference

[1] Fan D P, Gong C, Cao Y, et al. Enhanced-alignment measure for binary foreground map evaluation[J]. arXiv preprint arXiv:1805.10421, 2018.

[2] Fan D P, Cheng M M, Liu Y, et al. Structure-measure: A new way to evaluate foreground maps[C]. Proceedings of the IEEE international conference on computer vision. 2017: 4548-4557.