PAVI-net: pose-aware visual interaction for sports action quality assessment

Action quality assessment aims to evaluate how well an action is performed and has important potential applications in sports training and instructional feedback. However, video-based assessment methods may be distracted by complex backgrounds and may insufficiently capture fine-grained body configurations that are critical for distinguishing subtle differences in movement execution. To address these limitations, this study proposes a Pose-Aware Visual Interaction Network, termed PAVI-Net, for sports action quality assessment. PAVI-Net adopts a dual-stream architecture in which an RGB video stream captures global appearance and spatiotemporal motion, while a pose stream models body-joint configurations and motion dynamics. A pose-aware visual interaction module is introduced to use pose features as queries and video features as keys and values, enabling pose-guided refinement of visual representations through cross-attention. Unlike direct feature concatenation, the proposed interaction allows body-structure information to actively guide the selection of action-relevant visual features and reduce the influence of unrelated contextual regions. Experiments on the MTL-AQA and AQA-7 benchmark datasets demonstrate that PAVI-Net achieves a Spearman correlation of 0.9536 on MTL-AQA under the difficulty-degree setting and an average correlation of 0.8421 on AQA-7. Ablation results further show that pose-aware visual interaction improves performance over both the video-only baseline and direct video–pose concatenation. Nevertheless, the framework remains dependent on pose-estimation quality, and its performance may decrease under severe occlusion, rapid motion, low human visibility, or substantial domain differences. Therefore, the present results demonstrate the effectiveness of PAVI-Net on benchmark sports videos, while further validation is required before deployment in authentic sports teaching environments.