A methodology, metric, and use case for fast encoding screening strategies in video coding

The growing complexity of video encoding standards, particularly Versatile Video Coding (VVC), requires fast encoding techniques, often based on machine learning (ML), to enable practical encoder and decoder implementation. Traditional evaluation methods demand extensive computational resources, often resulting in hundreds of hours of encoding time, especially when assessing many models. This article proposes a novel methodology to systematically evaluate and compare early-stage ML-based approaches for accelerating video encoders, enabling assessments in reasonable timeframes while ensuring consistent comparisons. Additionally, the BD-rate-Time Reduction Ratio (BTR) metric is introduced to compare acceleration algorithms with a single value. A case study applies several ML classifiers to accelerate the first-level quadtree splitting in VVC, achieving a 20.18% time reduction with only a 0.96% BD-rate increase. Results show that the methodology supports large-scale model evaluation and improves the reliability of comparisons for future advances in video encoding efficiency.