Скачать или смотреть Anne Aaron, Netflix - Measuring Perceptual Video Quality at Scale [DEMUXED 2016]

We encode movies and TV shows for millions of viewers worldwide and aim to deliver the best video quality possible no matter what the content, viewing device or bandwidth. This effort hinges on one fundamental premise: that we can accurately and efficiently measure the perceptual quality of a video stream at scale. Simple metrics like PSNR and SSIM do not consistently reflect human perception while large-scale visual subjective testing is practically infeasible.

We developed Video Multimethod Assessment Fusion or VMAF as a method to more accurately measure video quality, and we’ve open-sourced the software. Since this tool applies machine-learning to fuse multiple metrics and features, it is easily extensible by integrating new metrics or adding new subjective training data that can improve the metric. We have used it to evaluate video codecs and encoding strategies. We are in the early stages of using VMAF as part of our automated QC and as a client-side metric to monitor system-wide A/B tests.