Sports motion scoring algorithms using visual transformers

Traditional visual recognition methods, although able to perform action classification, are difficult to achieve high accuracy because they insufficiently capture temporal dependencies, local deviations, and the relationship between continuous scores and discrete grades. To address these issues, this study presents a Visual Transformer model for sports motion scoring. Based on the Visual Transformer framework, the proposed algorithm jointly models movement normativity, completion, stability, and rhythm consistency through multi-dimensional score regression and grade classification. A dataset containing 3160 original sports-motion video samples was divided into training, validation, and test partitions at the athlete or source-video level. Performance was evaluated only on the held-out test set. The proposed model achieved an RMSE of 3.49, an MAE of 2.72, a Grade Accuracy of 84.0%, and a Spearman ρ of 0.908. These results indicate scoring potential under the current dataset, annotation protocol, and RGB-video setting, but do not establish cross-dataset or cross-sport generalizability. Compared with the corresponding ablation variants without TDEM, LDAM, and JOL, the full model reduces RMSE by 0.79, 0.92, and 0.48 points and improves Grade Accuracy by 5.7, 7.1, and 3.5 percentage points, respectively. These comparisons are internal ablation results obtained under the same dataset split and evaluation protocol. Under the same self-constructed dataset split and local reimplementation protocol, the proposed model achieved modest improvements over the included action-recognition and video-Transformer baselines, with an overall RMSE of 3.49, Grade Accuracy of 84.0%, and Spearman ρ of 0.908. These results support its comparative effectiveness within the present dataset and evaluation protocol, but they do not establish generalizability to independent datasets, unobserved sport categories, or different acquisition environments.