Home

Awesome

Adversarial Learning for Semi-supervised Semantic Segmentation

This repo is the pytorch implementation of the following paper:

Adversarial Learning for Semi-supervised Semantic Segmentation <br/> Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang <br/> Proceedings of the British Machine Vision Conference (BMVC), 2018.

Contact: Wei-Chih Hung (whung8 at ucmerced dot edu)

The code are heavily borrowed from a pytorch DeepLab implementation (Link). The baseline model is DeepLabv2-Resnet101 without multiscale training and CRF post processing, which yields meanIOU 73.6% on the VOC2012 validation set.

Please cite our paper if you find it useful for your research.

@inproceedings{Hung_semiseg_2018,
  author = {W.-C. Hung and Y.-H. Tsai and Y.-T. Liou and Y.-Y. Lin and M.-H. Yang},
  booktitle = {Proceedings of the British Machine Vision Conference (BMVC)},
  title = {Adversarial Learning for Semi-supervised Semantic Segmentation},
  year = {2018}
}

Prerequisite

Installation

git clone https://github.com/hfslyc/AdvSemiSeg.git
AdvSemiSeg/dataset/VOC2012/JPEGImages
                          /SegmentationClassAug

Testing on VOC2012 validation set with pretrained models

python evaluate_voc.py --pretrained-model semi0.125 --save-dir results

It will download the pretrained model with 1/8 training data and evaluate on the VOC2012 val set. The colorized images will be saved in results/ and the detailed class IOU will be saved in results/result.txt. The mean IOU should be around 68.8%.

Example visualization results

Training on VOC2012

python train.py --snapshot-dir snapshots \
                --partial-data 0.125 \
                --num-steps 20000 \
                --lambda-adv-pred 0.01 \
                --lambda-semi 0.1 --semi-start 5000 --mask-T 0.2

The parameters correspond to those in Table 5 of the paper.

To evaluate trained model, execute the following:

python evaluate_voc.py --restore-from snapshots/VOC_20000.pth \
                       --save-dir results

Changelog