3-Dimensional Convolution Based Iterative Model for Efficient Motion Map Generation for Representing Video Discriminative Information

Abstract

In this paper, we present a simple method to integrate the discriminative information of video for the action recognition tasks. We introduce the concept of motion map to represent the prefix of video sequences by optimizing the recognition accuracy of original video. 3-dimensional convolution (3Dconv) based model is used to generate the new motion map by integrating current motion map and future video frame. This model is capable of increasing the length of training video in iterative manner and allow us to generate the final motion map. Experimental evaluation results on widely used datasets i.e HMDB51 and UCF101 have revealed effectiveness and flexibility of proposed method over other baseline schemes.