Abstract
Streaming video accounts for a large fraction of the overall volume of mobile video data traffic. Because of network and throughput limitations, streaming video data may be afflicted by two dominant impairments: rebuffering events (where the video "freezes" - Fig. 1) and compression artifacts. Given that the end goal of every content provider is to maximize every end-user's visual quality of experience (QoE), subjective modelling of streaming video QoE has become a major concern. We have recently created a database to study the effects of temporal impairments on the behavior of human subjective raters. The new video QoE database contains long video sequences that were QoE-rated by human subjects under mobile viewing conditions, compressed to typical low bitrate values and subjected to realistic network and buffer constraints (Fig. 2). We observed that rebuffering was always obvious and unpleasant to subjects (Fig. 3) while bitrate (compression) changes tended to be less obvious due to content-related dependencies. On more compressible contents, transient bitrate drops (to avoid rebuffering) were preferred over rebuffering events, while consistently low bitrates were poorly tolerated. Further, long playback interruptions on higher quality videos led to larger drops in subjective QoE than on lower quality videos (Fig. 4). We also analyzed long and short term memory effects: when the perceived video quality was relatively stable, recent experiences were more influential (the "recency" effect) while impairments that occured early in a video activated longer term memory reactions (the "primacy" effect) (Fig. 5). We also evaluated a variety of high-performance objective video quality assessment algorithms on the new database and observed that they were unreliable predictors of visual QoE on videos that suffered from both rebuffering events and bitrate changes (Table I). We conclude that more general QoE models are needed that account for distortions, rebuffering events, and memory.
Meeting abstract presented at VSS 2017