Mining web videos for video quality assessment

Correlating estimates of objective measures related to the presence of different coding artifacts with the quality of video as perceived by human observers is a non-trivial task. There is no shortage of data to learn from, thanks to the Internet and web-sites such as YouTube^tm. There has, however, been little done in the research community to try to use such resources to advance our understanding of perceived video quality. The problem is the fact that it is not easy to obtain the Mean Opinion Score (MOS), a standard measure of the perceived video quality, for more than a handful of videos. The paper presents an approach to determining the quality of a relatively large number of videos obtained randomly from YouTube^tm. Several measures related to motion, saliency and coding artifacts are calculated for the frames of the video. Programmable graphics hardware is used to perform clustering: first, to create an artifacts-related signature of each video; then, to cluster the videos according to their signatures. To obtain an estimate for the video quality, MOS is obtained for representative videos, closest to the cluster centers. This is then used as an estimate of the quality of all other videos in the cluster. Results based on 2,107 videos containing some 90,000,000 frames are presented in the paper.