VMAF and variants: towards a unified VQA

Abstract

Video quality assessment (VQA) is now a fastgrowing subject, beginning to mature in the full reference (FR) case, while the burgeoning no reference (NR) case remains challenging. We investigate variants of the popular VMAF video quality assessment algorithm for the FR case, using support vector regression and feedforward neural networks, and extend it to the NR case, using the same learning architectures, to develop a partially unified framework for VQA. When heavily trained, algorithms such as VMAF perform well on test datasets, with 90%+ match; but predicting performance in the wild is better done by training/testing from scratch, as we do. Even from scratch, we achieve 90%+ performance in FR, with gains over VMAF. And we greatly reduce complexity vs. leading recent NR algorithms, VIDEVAL, RAPIQUE, yet exceed 80% in SRCC. In our preliminary testing, we find the improvements in trainability, while also constraining computational complexity, as quite encouraging, suggesting further study and analysis.