# Fully Convolutional Networks for Semantic Segmentation

PDF
https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
Author
Evan Shelhamer, Jonathan Long, and Trevor Darrell

## What is this?

• They trained CNN to train end-to-end and pixels-to-pixels and improved semantic segmentation.

## What is better than before?

• No need to pre- and post- processing such as superpixels
• Beter performance

## How it works?

• Each layer of convnet is $h \times w \times d$, and all layer computes a nonlinear filter.
• We can construe that FCN is convolutionalized typical CNN, which produces non-spatial outputs.
• To connect coarse outputs of CNN to dense pixels, we must upsample the output. They used deconvolution, which backwards convolution. It is easy to implement and possible to train by backpropagation.
• The loss is per-pixel multinomial logistic loss.
• They used pre-trained networks such as VGG16 and used transfer learning. They discarded final loss layer. Ignoring final pooling contributes to a better result.