Class Attention Transfer for Semantic Segmentation

Yubin Cho, Sukju Kang

Abstract: This paper proposes a knowledge transfer method using class attention maps (CAM) that are class-discriminative for training lightweight semantic segmentation networks. Since semantic segmentation classifies for each pixel, it is difficult to focus on the discriminative regions for each class in a single channel attention map. Thus, we generate attention maps for each class by using weights obtained from feature maps and class masks for squeezing the channels of the feature maps and then forcing a student network to generate the CAM that mimic the CAM of a teacher network. Our proposed method improves the state-of-the-art HRNetV2-W18+OCR by 4.78% in mIoU on the Cityscapes dataset.