Cross-channel Communication Networks

Cross-channel Communication Networks

2019-12-13 14:17:18

Paper: https://papers.nips.cc/paper/8411-cross-channel-communication-networks.pdf

Code: https://github.com/jwyang/C3Net.pytorch

SENet (Squeeze-and-Excitation Networks): http://openaccess.thecvf.com/content_cvpr_2018/papers/Hu_Squeeze-and-Excitation_Networks_CVPR_2018_paper.pdf

Code: https://github.com/moskomule/senet.pytorch

This paper introduces the graph neural network (GAT used in the experiments) into the regular CNN models to boost the interactions between different feature maps. Specifically speaking, it first utilize the CNN to extract the feature maps of the input, then, they reshape the feature maps into multiple vectors as the node of the graph. Then, they compute the similarity between different node the construct the adjacency matrix. The feature vectors and matix are inputted into the GAT module to conduct interactive learning. After that, they reshape the processed feature vectors into corresponding feature map, and feed them into the regular CNN modules.

Their experimental results demonstrate the effectiveness of this module in multiple CV tasks.