Long-term quality evaluation in OTT video

Abstract

Quality variations due to network bandwidth fluctuations is a common phenomenon in today's HTTP based over-the-top (OTT) video streaming services where video sessions often last several minutes. In addition to quality variations, OTT video playouts encounter buffer under-runs resulting in client-side video interruptions (a.k.a. stallings) which highly impact viewer's quality of experience (QoE). OTT service providers would like to know the overall quality score of a given video stream to optimize their service. Network provider's target is to maximize the overall video quality score for their network. In any case an accurate measurement of video stream quality is a key requirement for different stake holders in the OTT business. This paper investigates the design of a parametric model for estimating long-term (60 sec or longer) perceptual quality from short-term (10 sec) video quality estimates in H.264 based adaptive video streaming. Based on an advanced subjective test framework, the proposed model linearly combines the perceptual impact of three types of distortions (initial-loading, stalling and coding/quality-switching) observed in OTT video in an overall quality score on a 5-point opinion scale. The proposed model yields high accuracy of quality prediction when evaluated for up-to 3 minutes long video sequences.