Fast Versatile Video Coding using Specialised Decision Trees
Gosala Kulupana, Venkata Phani Kumar M, Saverio Blasi
Abstract
New video compression technology is being developed to meet the increasing demands for more content at higher resolutions. To this end, the Versatile Video Coding (VVC) standard was recently finalised by the Joint Video Experts Team (JVET). Even though VVC can considerably reduce the bitrate requirements of compressed videos, the complexity of encoders has also drastically increased. In order to allow practical implementations to meet complexity requirements, this paper proposes a new block partitioning scheme for VVC inter-coding which can significantly reduce the complexity of the encoding. The paper is based on a novel feature selection scheme to speed up the partitioning decisions in VTM. The features are fed into classifiers where both a new random forest tree selection process and an early termination criteria are applied to better specialise the predictions to block characteristics, and allowing uniform distribution of complexity savings throughout the sequence. The proposed method was shown to consistently reduce complexity of VVC inter-coding with marginal increases in bitrate.