The 11th Iranian and the first International Conference on Machine Vision and Image Processing

Class Attention Map Distillation for Efficient Semantic Segmentation

Nader Karimi Bavandpour, Shohreh Kasaei
The 11th Iranian and the first International Conference on Machine Vision and Image Processing (MVIP 2020)

Abstract

In this paper, we introduce a novel way of capturing the information of a powerful and trained deep convolutional neural network and distill it to a training smaller network. Our method, despite of many others which work on final layers, can successfully extract suitable information for distillation from intermediate layers of a network by making class specific attention maps and then forcing the student network to mimic those attentions. We apply this method to state of the art semantic segmentation architectures and show its effectiveness by experiments on the standard Pascal Voc 2012 dataset.

Keywords: Semantic Segmentation, Knowledge Distillation, Saliency Maps



© 2017-2021 ISMVIP All Rights Reserved