Fully Convolutional Network with CRF postprocessing in Pytorch

Pytorch based implementation of Fully Convolutional Networks for Semantic Segmentation CVPR'15 paper.

Requirements

1. pytorch >=0.3
2. pydensecrf
3. opencv>3.0
4. matplotlib
5. numpy
6. scikit-image
7. scikit-learn

Download the dataset from this link and place it in the main directory: https://github.com/mostafaizz/camvid .

Dataset contains the following files:

1. 701_StillsRaw_full
2. LabelApproved_full
3. label_colors.txt
4. label_colorsSorted.txt

Preprocessing

python camvid_utils.py

It will create a folder in the directory which contains the above folders,a directory named Labeled_idx which would contain the 32 label encoded numpy vectors.

Training

python camvid_train.py

Metric: Average IOU

Training Results:

Coming Soon!!