Noh, Hyeonwoo, Seunghoon Hong, and Bohyung Han. “Learning deconvolution network for semantic segmentation.” Proceedings of the IEEE International Conference on Com-
puter Vision. 2015. (Citations: 139).
[1]. http://web.cs.hacettepe.edu.tr/ ̃aykut/classes/spring2016/bil722/slides/w06-deconvnet.pdf.
puter Vision. 2015. (Citations: 139).
1 Motivation
The deconvolution layer of FCN is too coarse and overly simple.
2 Architecture
See Fig. Region proposals are fed into the network. Compared with feeding the whole image into the network, it can handle object in various sizes.
The network is composed with a VGG followed by a upside down (mirrored) VGG. The forward pass of unpooling layer is the backward pass of standard pooling layer. The
unpooling switches are from the correponding pooling layer.
[1]. http://web.cs.hacettepe.edu.tr/ ̃aykut/classes/spring2016/bil722/slides/w06-deconvnet.pdf.