Learning-Based Fast VVC Affine Motion Estimation
Fernando Sagrilo, Marta Loose, Ramiro Viana, Gustavo Sanchez, Guilherme Corrêa, Luciano Agostini
Abstract
This paper presents a fast Affine Motion Estimation (AME) of Versatile Video Coding (VVC) Standard, based on Machine Learning and using Random Forest (RF) classification method. This encoding approach develops an RF model for each block size. The models were trained with information extracted during the VVC encoding process of the current, parent, and neighboring Coding Units (CU). Each model is applied to predict whether the Affine Motion Estimation (AME) will be skipped or not for that CU size. The proposed solution achieves a reduction of 20% on average in AME encoding time, with an insignificant impact of 0.07% on BD-BR.
Topics & Concepts
Affine transformationCoding (social sciences)Motion estimationComputer scienceEncoding (memory)Artificial intelligenceRandom forestRandom accessAffine shape adaptationBlock (permutation group theory)Computer visionBlock sizeMotion (physics)Pattern recognition (psychology)AlgorithmMathematicsAffine combinationStatisticsOperating systemKey (lock)GeometryComputer securityPure mathematicsVideo Coding and Compression TechnologiesAdvanced Image Processing TechniquesAdvanced Data Compression Techniques