Between Two and Six? Towards Correct Estimation of JND Step Sizes for VMAF-based Bitrate Laddering

We currently witness the rapidly growing importance of intelligent video streaming quality optimization and reduction of video delivery costs. Per-Title encoding, in contrast to a fixed bitrate ladder, shows significant promise to deliver higher quality video streams by addressing the trade-off between compression efficiency and video characteristics such as resolution and frame rate. Selecting encodings with noticeable quality differences in between prevents the construction of an inefficient bitrate ladder that suffers from too similar quality encodings. In this respect, the VMAF metric represents a promising foundation for bitrate laddering, as it currently yields the highest video quality prediction performance. However, the minimum noticeable quality difference, referred as to just-noticeable-difference (JND), has not been properly validated for VMAF yet, with existing sources proposing highly diverse ΔVMAF step sizes ranging from two to six. This paper investigates JND for VMAF to facilitate the efficient construction of content-specific bitrate ladders. Using a publicly available JND multimedia dataset, we invalidate existing step size rules of thumb by revealing the large errors they introduce when applied to real-world video content. We investigate the different factors that influence the actual ΔVMAF step sizes required to obtain correct JNDs and develop a model using content features that more accurately determines the optimal quality step sizes required for properly laddering individual clips.