Gated Feedback Refinement Network for Dense Image Labeling
This paper is a follow up of Label Refinement Network. The model is based on a encoding-decoding architecture with skip connections, the encoding is done by a VGG-16 network where only the convolutional layers are used. The novelty of their work is the use of a gating mechanism in the skip connection.
As you can see they use multiple crossentropy losses for each resolution of feature maps on the decoding part. This help the network to extract better information at each resolution.
Gate Refinement unit
The Gate unit merges two resolution of feature maps by upsampling the lower resolution and make an element wise product of the features. These features are finally handled by the Refinement unit that produces a label segmentation map.
The main idea behind this is that lower resolution label maps integrate higher-frequency details from the skip connections and that lead to better segmentation maps.
They test their method on three datasets namely, CamVid, Pascal VOC 2012, and a subset of Pascal VOC 2012 with only horses and cows where the task is to separate each part of the animal.
PASCAL VOC 2012