Context-Aware Memory Attention Network for Video-Based Action Recognition

Human action recognition is very popularly researched in the computer vision community. The current challenge is to render it efficient enough for wide deployment. In this paper, we propose a human action recognition model which does not require optical flow extraction and 3D convolution, called Context-Aware Memory Attention Network (CAMA-Net). It consists of an attention module called Context-Aware Memory Attention Module which is used to calculate the relevance score between the key and value pairs from the backbone output. The proposed method is evaluated and tested on popular public action recognition datasets, UCF101 and HMDB51. The results demonstrate the strength of our proposed model as it outperforms existing baseline models.