Action Recognition by 3D Convolutional Network

Abstract

In this paper we focused on sport action recognition with 3D convolutional network. We executed two different experiments. In the first experiment we distinguished two similar activities: running and walking. The experiment proved that 3D convolutional network is able to learn spatio-temporal features of video sequence. We compared three different 3D convolutional networks and the best accuracy was achieved by architecture Al reached 85%. The second experiment was performed on subset of UCF101 dataset. We selected 15 activities. Architecture A1 achieved accuracy 80.7%. Our results show that it is possible to achieve relatively high accuracy using subtile architecture of 3D convolutional network.