Quality assessment of sports actions based on adaptive-UniFormer

Abstract

Sports action quality assessment (AQA) presents significant challenging, requiring comprehensive evaluation of motion completeness, fluency, and difficulty level for accurate quality scoring. This paper proposes Adaptive-UniFormer, an innovative AQA network integrating an Adaptive Token Halting Mechanism (ATHM) based on the UniFormerV2 architecture. The framework introduces the Top-K selection mechanism in local feature extraction to efficiently eliminate redundant background tokens, and ATHM in the global feature extraction to focus computation on action-related tokens, significantly reducing computational overhead. Final action classification and quality scores are generated through multi-stage feature fusion and a Multi-Layer Perceptron (MLP). Comprehensive experiments demonstrate superior performance, for action recognition, the model achieves 87.6 % Top-1 and 98.7 % Top-5 accuracy on the UCF101, while reducing computational costs by 46.5 % in FLOPs, along with 78.4 % Top-1 accuracy On HMDB51. For action quality assessment, it obtains average Spearman’s rank correlation coefficient of 0.8223 on AQA-7 and 0.9502 on MTL-AQA. In conclusion, the proposed Adaptive-UniFormer establishes new benchmarks for recognition accuracy, computational efficiency, and AQA performance, offering an effective solution for sports action analysis.