Some differences from original FCN

[zengxianyu]

Hi, I'm trying to use FCN in pytorch, and your work interests me very much. It really helps me a lot, but I found the segmentation results are not so satisfying (after 30 epoch of training on pascal2012), and there are some differences from original FCN implementation.

• in the original FCN8s, the outputs of pool3, pool4, and pool5 are sent into subsequent convolution layers to produce segmentation results, however, in your version, it's the outputs of conv3, conv4, and conv5 that are sent into the subsequent convolution layers!
• in the original FCN8s implementation, the output of pool4 (and it's the same with pool3) is multiplied with a small constant, before being sent into the subsequent convolution layers to produce segmentation maps.
• in the original FCN8s implementation, the VGG convolution layers are also updated during training. But in your version, VGG convolution layers won't be updated.
• the original implementation uses deconvolution to upsample.

[Bodo Kaiser]

Hi, you have very good arguments!

Unfortunately I have very limited computing resources so I cannot update this project at the moment
but feel free to send a PR with your changes (if you are interested).