[深度学习论文笔记][Semantic Segmentation] Fully Convolutional Networks for Semantic Segmentation

Long, Jonathan, Evan Shelhamer, and Trevor Darrell. “Fully convolutional networks for semantic segmentation.” Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition. 2015. (Citations: 863).

1 Motivation

Previous works use interpolation to up-sample the output predictions to the same size as the input. Instead, we let the model learn the up-sampling weights.

2 Pipeline

See Fig. Fully convolutional networks are used to take input of arbitrary size and produce correspondingly-sized output. Feature maps are up-sampled using deconvolution.

The forward pass of deconvolution is the same as backward pass of standard convolution.

3 Skip Connections
See Fig. We use skip connections to combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate
and detailed segmentations. See results in Fig.

4 References
[1]. http://techtalks.tv/talks/fully-convolutional-networks-for-semantic-segmentation/61606/.