3D CNNs for Pose Classification Using KB Dataset

Abstract

Human action prediction and classification is a hot area of research that is employed in various fields. Many of 2D action classification algorithms for images using CNN are developed. But, if we try to apply them on fresh datasets like KB dataset we get an average accuracy prediction due to occlusionand foreshortening of limbs, different orientation and rotation. In this paper, we employed a 3D deep convolutional neural network for human action classification. We capture various stillsof the few  actionsin different views, arrange them together like frames in video and trained a 3D CNN. This gives network the ability to leverage various context of the action that lead to an improved performance and accurate classification of poses by the network.