Skip to content
/ SGAN Public

Saliency Guided Self-attention Network for Weakly and Semi-supervised Semantic Segmentation(IEEE ACCESS)

License

Notifications You must be signed in to change notification settings

yaoqi-zd/SGAN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Saliency Guided Self-attention Network for Weakly and Semi-supervised Semantic Segmentation (IEEE ACCESS)

This code is a implementation of the high-quality seed generation step in the paper SGAN. The code is developed based on the Pytorch framework.

PWC PWC

Introduction

Overview of SGAN The proposed approach consists of three components: (1) a CNN backbone to learn deep feature representations; (2) a saliency guided self-attention module that propagates attentions from small discriminative parts to non-discriminative regions; (3) an image classification branch and a seed segmentation branch to supervise the training of the entire network.

Installation

1. Prerequisites

Tested on Ubuntu16.04, CUDA9.0, python3.5, Pytorch 0.4.1, CRF, NVIDIA RTX 2080TI

2. Dataset

  1. Download the prepared VOC2012 dataset, it contains all the files to reproduce the paper's results, including training images, pre-computed saliency maps and initial seeds.
  2. modify the "dataset_root" variable in config file correspondingly.

3. models

  1. Download vgg_init.pth model for initializing the SGAN network.
  2. You can also download the trained SGAN model model_iter_8000.pth for seed generation.

Usage

1. train SGAN network

CUDA_VISIBLE_DEVICES=0,1 python tools/stage2/train_sgan.py  --cfg_file config/sgan_vgg16_321x321.yml

2. generate high-quality seeds

CUDA_VISIBLE_DEVICES=0 python tools/stage2/infer_cam.py --cfg_file config/sgan_vgg16_321x321.yml

3. (optionally) evaluate the seed quality

python tools/eval_mIoU.py --res_path [seed_path] --num_classes 22

Seed Result

If you have finished all the steps above, you will get the seeds with precision=76.38 recall=57.26. The result is slightly different with that reported in our paper(precision=76.37, recall=57.35) since this repo is seperated from a larger project and the code may be slighted modified.

The pre-computed seeds can be downloaded here.

Segmentation Result

If you want to reproduce our segmentation results, please refer DSRG for experiment setup, you may need to compile caffe, setup caffe python path and some other path variables. We provide the dsrg code, caffe with modified data layer in dsrg folder. In experiment/anti_noise folder, run

./run_anti_noise.sh [gpu_id]

Note that the our modified dsrg code is a little dirty and I don't have time to clean up. However, it's enough to reproduce the segmentation performance.

License

SGAN is released under the MIT license

Citing SGAN

if you find SGAN useful in your research, please consider citing:

@article{yao2020saliency,
  title={Saliency Guided Self-Attention Network for Weakly and Semi-Supervised Semantic Segmentation},
  author={Yao, Qi and Gong, Xiaojin},
  journal={IEEE Access},
  volume={8},
  pages={14413--14423},
  year={2020},
  publisher={IEEE}
}

if you also use the SNet to generate saliency maps, please consider citing:

@article{xiao2018deep,
  title={Deep salient object detection with dense connections and distraction diagnosis},
  author={Xiao, Huaxin and Feng, Jiashi and Wei, Yunchao and Zhang, Maojun and Yan, Shuicheng},
  journal={IEEE Transactions on Multimedia},
  volume={20},
  number={12},
  pages={3239--3251},
  year={2018},
  publisher={IEEE}
}
@inproceedings{huang2018dsrg,
    title={Weakly-Supervised Semantic Segmentation Network with Deep Seeded Region Growing},
    author={Huang, Zilong and Wang, Xinggang and Wang, Jiasi and Liu, Wenyu and Wang, Jingdong},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    pages={7014--7023},
    year={2018}
}

About

Saliency Guided Self-attention Network for Weakly and Semi-supervised Semantic Segmentation(IEEE ACCESS)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published