search

Boyan-Lenin-Xu / GCResNet

PyTorch implementation of Graph Convolutional Networks in Feature Space for Image Deblurring and Super-resolution, IJCNN 2021.
12 stars 0 forks source link

a little question #1

Open PASSENGER128 opened 3 years ago

PASSENGER128 commented 3 years ago

Hello, this is an excellent job. When I was reading your paper, I had a question: You do graph convolution network on the feature space and it feels similar to channel attention. Is there any difference between them? In other words, how is doing Graph convolution in feature space better than channel attention?

I'm looking forward to your answer.Thank you.

Boyan-Lenin-Xu commented 3 years ago

Hi, thank you for your attention. Graph structure can give a detailed consideration about different importance of different features. For instance, a node in a graph could be very important/important/less important/no importance (because the degree of each node is diverse, e.g., 3,4,6,8...). While channel attention can only select some features to focus on. In addition, the relation between node can also be better expressed by graph structure.