An Empirical Study for Efficient Video Quality Assessment

Abstract

Video quality assessment (VQA) plays a critical role in optimizing various video processing systems, yet achieving both high accuracy and computational efficiency remains a challenging task. Recent advances in deep neural network (DNN)-based VQA models have led to notable performance improvements, but often at the cost of high computational complexity and memory consumption, limiting their applicability in resource-constrained scenarios. In this paper, we empirically investigate a set of good practices for building efficient yet effective VQA models. Specifically, we decompose the VQA training pipeline into three components: video preprocessing, quality-aware feature extraction, and optimization techniques. For each component, we identify and validate effective practices using the KVQ dataset---a user-generated content (UGC) VQA dataset that includes both in-the-wild distortions and processing-induced artifacts such as compression and enhancement. Based on these findings, we propose E-VQA, an efficient VQA model that combines the best-performing practices. Experiments conducted on the KVQ dataset, as well as the large-scale UGC VQA dataset LSVQ, demonstrate that E-VQA achieves competitive performance while significantly reducing computational complexity. Furthermore, E-VQA ranked third in the NTIRE 2025 Short-form UGC Video Quality Assessment Challenge, highlighting its practical effectiveness. The code is available at https://github.com/sunwei925/E-VQA.