Multi-preset video encoder bitrate ladder prediction
Fatemeh Nasiri, Wassim Hamidouche, Luce Morin, Nicolas Dhollande, Jean-Yves Aubié · 2022
Determining optimal resolution for any possible combination of video title and subscriber bandwidth limit is a critical problem to address in streaming and Video on Demand (VoD) services. Approaches such as per-title encoding aim at addressing this problem through computationally complex encoding passes for constructing content-adaptive bitrate ladders. This paper proposes a machine learning (ML)-based method that accelerates such approaches by estimating bitrate ladders of a multi-preset encoder. To do so, reference content-adaptive bitrate ladder using the fastest possible preset of a given encoder is constructed exhaustively. Then, an offline-trained regressor is deployed to transform the fast-preset ladder into the slow-preset ladder that is needed for actual encoding pass. The experiment, carried out on the official optimized codec implementation of the Versatile Video Coding (VVC) standard (VVenC), shows a significant performance improvement in terms of compression efficiency and complexity trade-off. In particular, the proposed method can offer 91% ladder construction complexity reduction at the cost of only 0.88% coding efficiency drop in terms of BD-BR.