Cognitive Calibration Meets Topological Graph Learning for Fine-Grained Sports Action Assessment

Abstract

Fine grained sports action evaluation not only requires identifying action categories, but also quantitative analysis of execution quality, technical standardization, and even subtle deviations. This has put forward higher requirements for the granularity of feature extraction, the accuracy of temporal modeling, and the uncertainty awareness ability of the model. This article proposes an action evaluation model that combines cognitive calibration mechanism with multi granularity topology learning. We named it Cognitive Calibration Topology Graph Convolutional Network (CCT-GCN). This model constructs a dynamic enhanced graph structure, whose edge set consists of structural edges naturally connected by human bones and functional edges dynamically generated based on kinematic correlations. The functional edge is obtained by calculating the correlation coefficient of the node motion feature sequence within the time window. To construct a multi granularity receptive field, the model adopts an iterative coarsening strategy based on differences in exercise intensity. The theoretical basis of this strategy is to perform hierarchical merging of nodes, generate graph sequences with decreasing scales, and perform graph convolution operations in parallel. Introduce multi-scale time window convolution in the temporal dimension. This convolution captures fine motion features of different time spans through the division of different window scales and step sizes. The cognitive calibration module first models a multivariate Gaussian distribution in the high-order feature space, using the trace of the covariance matrix as a measure of cognitive uncertainty. Subsequently, a soft calibration method based on self attention mechanism is constructed, and the variance information of the feature components is injected into the attention weight calculation. Experiments on three benchmark datasets, FineDiving, MTL-AQA, and UNLV-DVS, showed that the Spearman correlation coefficients of CCT-GCN reached 0.8273, 0.9451, and 0.8913, respectively. The ablation experiment confirmed that the multi granularity topology module alone contributed 3.34 percentage points of performance improvement. Robustness analysis shows that under the condition of 30% keypoint occlusion, the correlation coefficient of CCT-GCN only decreases by 6.36%, while the variant of ablation calibration module decreases by 13.70%.