International Conference on Computer and Knowledge Engineering

Home / 12th International Conference on Computer and Knowledge Engineering

Attention Transfer in Self-Regulated Networks for Recognizing Human Actions from Still Images

Authors :

Masoumeh Chapariniya¹ Sara Vesali Barazande² Seyed Sajad Ashrafi³ Shahriar B.Shokouhi⁴

1- School of Electrical Engineering, Iran University of Science and Technology 2- School of Electrical Engineering, Iran University of Science and Technology 3- School of Electrical Engineering, Iran University of Science and Technology 4- School of Electrical Engineering, Iran University of Science and Technology

Keywords :

Attention transfer،human action recognition; self-regulated networks; transfer learning.،human action recognition،self-regulated networks،transfer learning

Abstract :

Human action recognition in still images (HAR) is a challenging computer vision task owing to problems such as lack of temporal information and large intra-class variations, cluttered backgrounds, and misleading objects which requires highly discriminative features. Transfer learning algorithms such as knowledge distillation and attention transfer techniques offer the necessary abilities in producing informative features by preserving prior knowledge while learning new representations. Because The ResNet and its variants have made significant advances in computer vision, most research to date focused on knowledge distillation and attention transfer in this architecture. Recently, self-regulated networks based on regulator module have been introduced that perform better than ResNet networks in various computer vision tasks. In this article, we propose the attention transfer framework in self-regulated networks for human action recognition in still images. We conduct extensive experiments on Stanford 40 and Pascal Voc 2012 Action datasets to investigate the performance of the proposed framework. The best setting of our method gains 93.17% (in terms of mAP) on Stanford40 dataset and 91.83% (in terms of mAP) on Pascal Voc 2012 Action datasets. Experiments demonstrate that attention transfer framework in self-regulated networks with extraction more representative and informative features through regulator module based on memory mechanism and without using any auxiliary data such as personal bounding box, objects bounding boxes, and human-object interactions has been able to significantly improve the action recognition in still images.