Contextual Action Recognition in Videos using Tube Convolutional Neural Network

Abstract

In an image classification and object detection Deep learning has been exhibited to accomplish great results.But deep learning on video analysis has been limited due to complexity of video data and lack of annotations. In this paper,we propose Tube Convolutional Technique (T-CT) for action detection in videos. The proposed architecture is a unified deep network that is able to identify and localize action based on 3D convolution features. A video is first divided into equal length eight frame clips and next for each clip a set of tube proposals are generated based on 3D TCT features. Finally, the tube proposals of differents are coupled along using network flow and spatio-temporal action detection is performed victimisation these linked video proposals.