Knowledge Distillation Framework for Action Recognition in Still Images
Masoumeh Chapariniya, Seyed Sajad Ashrafi, Shahriar B. Shokouhi · 2020
Still image action recognition is generally challenging, despite the considerable progress of convolutional neural networks in image classification. The absence of motion cues in still images is a chief challenge in human action recognition from images. Currently, most efficient methods train a deep CNN network directly on action recognition images. However, these methods have many parameters and high computational costs. Furthermore, most available methods have extracted auxiliary data such as human body gestures, involved objects, and the appearance of body parts from images. These methods employ bounding boxes as auxiliary input in the training and testing phases. However, this solution is inefficient because bounding boxes are not always available. We propose a knowledge distillation technique for action recognition in still images. This technique distills the knowledge of the cumbersome model into a small model. This framework also eliminates the requirement of using bounding boxes during training and testing networks. The best mean Average Precision (mAP) of our method is 92.11% on Stanford40 dataset.