Benchmarking AIGC Video Quality Assessment: A Dataset and Unified Model

• DOI: 10.48550/arxiv.2407.21408
• Publication Date: 2024-07-31
• Authors: Zhichao Zhang; Xinyue Li; Wei Sun; Jun Jia; Xiongkuo Min; Zicheng Zhang; Chunyi Li; Zijian Chen; Puyi Wang; Zhongpeng Ji; Fengyu Sun; Shangling Jui; Guangtao Zhai
• Keywords: Benchmarking; Quality Assessment

In recent years, artificial intelligence (AI) driven video generation has garnered significant attention due to advancements in stable diffusion and large language model techniques. Thus, there is a great demand for accurate quality assessment (VQA) models measure the perceptual of AI-generated content (AIGC) videos as well optimize However, assessing AIGC quite challenging highly complex distortions they exhibit (e.g., unnatural action, irrational objects, etc.). Therefore, this paper, we try systemically investigate AIGC-VQA problem from both subjective objective perspectives. For perspective, construct Large-scale Generated Vdeo Quality (LGVQ) dataset, consisting 2,808 generated by 6 using 468 carefully selected text prompts. Unlike previous VQA experiments, evaluate three dimensions: spatial quality, temporal text-to-video alignment, which hold utmost importance current establish benchmark evaluating existing metrics on LGVQ reveals that perform poorly dataset. propose Unify Video (UGVQ) comprehensively accurately across aspects unified model, uses visual, textual motion features corresponding prompt, integrates key enhance feature expression. We hope our can promote development evaluation videos. The dataset UGVQ metric will be publicly released.