Attention Transfer in Self-Regulated Networks for Recognizing Human Actions from Still Images
Masoumeh Chapariniya, Sara Vesali Barazande, Seyed Sajad Ashrafi, Shahriar B. Shokouhi · 2022
Recognizing human actions from still images is a demanding computer vision task because of problems such as lack of temporal information, large within-class variations, cluttered backgrounds, and misleading objects that require highly discriminatory features. Transfer learning algorithms such as knowledge distillation and attentional transfer techniques provide the necessary skills to generate valuable features by maintaining former knowledge while new demonstrations are learned. Because the Residual networks have made significant advances in machine vision, most research to date focused on knowledge distillation and attention transfer in this architecture. But the self-regulated networks, which were recently introduced based on regulator modules and have better performance than ResNet networks in various computer vision tasks, have been less investigated. In order to recognize human actions in images, we propose the attention transfer framework in self-regulated networks. In addition to extracting representative features from images, the self-regulated module does not utilize annotations such as personal bounding boxes, object bounding boxes, and human-object interactions. To study the performance of our proposed method, extensive experiments were conducted on Stanford 40 and Pascal Voc 2012 datasets. Our best set of algorithms gains 93.17% on Stanford40 and 91.83% on Pascal Voc 2012. In this attention transfer framework, self-regulated networks are employed, which are based on the memory mechanism module.