A Dual Stream Model for Activity Recognition: Exploiting Residual- CNN with Transfer Learning

Abstract

Visual content has a protagonist role in this age of data revolution. These days, computer vision research community is fascinated towards application of convolution neural networks and transfer learning for various image and video analysis tasks. Residual connection in CNN can facilitate the training process in the deep networks. This paper investigates and uses deep residual networks with fusion based dual stream pre-trained models for activity recognition from video streams. The architecture is further trained and evaluated using standard video actions benchmarks of UCF-101, HMDB-51 and NTU RGB. Performance of depth-based variants of residual networks is also analysed. The proposed approach not only provides competitive results but also better at exploiting pre-trained model and annotated image data. © 2020 Informa UK Limited, trading as Taylor & Francis Group.