Action Recognition Using Temporal Partitioning of Motion Information

In this paper, a temporal representation method for video action recognition is proposed. Since the intrinsic property of the video stream is its temporal variation, the optical flow images are calculated to show the short-term motion. In order to avoid training a complex network from scratch, a pre-trained network is utilized for frame-level feature extraction. For video level representation, pyramidal pooled time series is considered since the short-term variation can be captured in order to represent fixed-size long-term features. In addition, to solve the information missing problem through long videos, a simple video level representation using temporal partitioning is proposed too. The experimental results of the proposed method illustrates an acceptable performance both in first and third-person action recognition.