For piwise I ported various models from caffe prototxt to pytorch including FCN, Segnet, UNet and PSPNet.
After training them about 30-40 epochs on the PASCAL VOC12 segmentation dataset I get following results:
So PSPNet, FCN8 just output black images. UNet, SegNet give some results but are still far away from the results claimed in their paper.
Links to networks and training.
Any suggestions what may be the problem or what I could try next?