Optimizing UVQ for scaling video quality assessment
Abstract
UVQ (Universal Video Quality) is an ML-based metric that offers state-of-the-art accuracy for no-reference video quality assessment. This work outlines the enhancements made to the UVQ-stack enabling quality assessment in a large scale. Our improvements span three complementary axes: system efficiency, model efficiency, and dataset representativeness. For system efficiency, we developed an FFmpeg filter that handles accurate frame sampling, decoding and scaling directly from video bytes before delivering frames to the UVQ model. This integration of decoding, preprocessing, and inference within a single filter eliminates the need for temporary disk writes, lowers memory usage, and ensures precise control over frame selection. We further boosted throughput by reshaping tensors to enable batch-processing of patches and avoiding expensive per-patch inference loops. For model efficiency, we distilled UVQ into a more lightweight architecture for each of its components. Features from these components are fed into a thin aggregation layer that predicts final quality scores. This approach reduces the compute from around 16 GMACs (UVQ v1.5) to 1.38 GMACs for a video of 1-sec duration, while maintaining high prediction performance, demonstrated by a 0.92 Spearman-correlation with the teacher model (UVQ v1.5) on our test set. Finally, we propose a practical data curation method for UVQ training and evaluation that enables us to maintain metric accuracy across diverse video content. Our method constructs representative datasets that balance content diversity, popularity-weighted sampling, and long-tail distribution. Together, these contributions enable UVQ-based quality assessment to scale across modern video ecosystems.