An Investigation of Automated Design of VMAF

Video quality assessment (VQA) methods have many applications in compression technologies, video transmission planning, and monitoring quality of service. Netflix has recently developed an open-source, state-of-the-art VQA method called video multi-methods assessment fusion (VMAF) which proved to correlate well with the human judgments of perceptual quality of videos. VMAF is designed manually by human experts. To our knowledge, no prior study has been conducted to automate the design of VMAF. In this paper, we propose a simulated annealing search (SA) and a genetic algorithm (GA) to search the space of VMAF configurations. We used the Netflix public dataset and the LIVE-VQA dataset to evaluate the proposed approaches. It is found that both SA and the GA improve the design of VMAF with the GA being the superior method.