DVD: Constructing a Discriminative Video Descriptor by Convolving Frame Features

The core to organize, classify, search, compare and retrieve videos is comparing the video descriptors. In this paper, we propose a Discriminative Video Descriptor (DVD) which is a general way to build the video descriptors on top of various frame features. We built the DVD on top of the HSV-color distribution and evaluated its performance for the Near-Duplicate Video Detection task by using the CC_WEB_VIDEOS dataset. The average detection accuracy achieved 94.4%. We also evaluated the DVD for Human Action Recognition task by building the DVD on top of the 3D-SIFT with Weizmann human action dataset. The average recognition accuracy achieved 97.84%. In practice, the DVD only introduce slightly computational overhead. The average time to build the DVD on top of the HSV-color distribution and 3D-SIFT for a single video was 0.128 $s$ (average 11 frames) and 0.04 $s$ (200 interest points), respectively.