Accepted by IEEE AICAS 2022 (Oral) [Paper]
We generate the class attention maps (CAM) from the feature maps of each student and the teacher network, and then, the student network is guided to mimic the regions that need to be paid more attention. The overall architecture of Class Attention Transfer is illustrated as below.
The CAM generated by our method from the feature map of the teacher network shows activated regions corresponding to the categories.