Fully Convolutional Networks[1] implemented in PyTorch. Although [2], [3] have implemeted it very well, the purpose of this repository is for me to gain familarity with semantic segmentation with pytorch implementations. Some of the code are borrowed from [2], [3], [4]. Thanks them very much.

• python 3.6
• pytorch==0.3.0, torchvision, scipy, ...

Note:

• All the code are developed and tested on Python 3.6 and maybe not support Python 2.x

• You can install all the python packages one-line by running:

sudo pip3 install -r requirements.txt

Support Pascal VOC 2012 dataset.

1. Download data by running:

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz

2. Extract the VOCtrainval_11-May-2012.tar and benchmark_RELEASE.tgz, modify the data path in config.py.

• To train the model:

python main.py --phase train

• To val the model:

python main.py --phase val

• To test

python main.py --phase test --in_path xxx/xxx.jpg --out_path ./results/

Displayed Raw jpg	Displayed Ground truth label	Displayed Predictions jpg

[1] Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015: 3431-3440.

[2] https://github.com/wkentaro/pytorch-fcn

[3] https://github.com/meetshah1995/pytorch-semseg

[4] https://github.com/chenyuntc/pytorch-best-practice