Genre-Aware Representation Learning for No-Reference Game Video Quality Assessment

The proliferation of game video streaming has created an urgent need for automated Video Quality Assessment (VQA) to ensure user Quality of Experience (QoE). However, gaming videos, being synthetically rendered, present a pronounced content heterogeneity in genre, art style, and motion dynamics that invalidates the assumptions of conventional VQA models designed for Natural Scene Content (NSC). This heterogeneity causes the perceptual impact of distortions to be highly context-dependent. Therefore, models designed for NSC will show significant performance degradation when applied to gaming content. More critically, many existing game-specific VQA methods still operate in a content-agnostic manner, failing to achieve robust performance across the diverse spectrum of game types. To address this critical gap, we propose MTLGVQA, a No-Reference VQA framework that leverages Multi-Task Learning. Our approach reframes quality assessment by jointly learning a primary quality regression task and an auxiliary genre classification task. These tasks are processed through a shared feature encoder and governed by an adaptive loss weighting strategy. This paradigm compels the model to learn a unified representation that is simultaneously sensitive to low-level distortion artifacts and high-level semantic context. Extensive experiments on public gaming VQA datasets demonstrate that MTLGVQA not only achieves state-of-the-art prediction accuracy but also exhibits superior generalization capabilities across disparate game genres.