Hi,
Thanks for sharing the code. I am little unsure about the dataset that shoud be used for training. Also, how it should be pre-processed. The paper says that PASCAL VOC 2012 has been used with some CRF pre-processing on the polygon annotations provided for the segmentation.

I have downloaded PASCAL VOC 2012 from http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#data . The images folder contains 17000 images. How should I generate the ground truth from this?

Hoping to get some response.

Thanks