AutoSco: automatic scoring framework for figure skating action based on C3D and attentive Bi-LSTM models

Current mainstream scoring methods perform poorly in segmenting action features in long videos, leading to poor robustness issues. Therefore, this study proposes a bidirectional long-short term memory model (Bi-LSTM) combining C3D and multi-layer attention mechanism. Specifically, C3D is responsible for extracting three-dimensional spatiotemporal features from figure skating videos. A multi-layer attention mechanism, inserted between the convolutional layers and the pooling layers, includes parallel feature and temporal attention channels: the former assigns weights to key action regions within each time step, while the latter emphasizes temporal correlations across consecutive steps, enabling the entire model to automatically focus on important features and temporal dimension information in the input features. Furthermore, the Bi-LSTM is introduced for fine-grained scoring of skating movements, improving interpretability by capturing contextual dependencies. Experimental results on FIS-V and FS1000 datasets demonstrate the superiority of the proposed AutoSco: on the FIS-V dataset, its TES accuracy exceeds that of STPE by 15 %, with a PCS Spearman Rank Correlation (SRC) of 75 %. On the FS1000 dataset, it achieves SRC values of 89 % (TES) and 85 % (PCS), with MSE values of 80 % (TES) and 60 % (PCS), which are lower than all baseline methods. These results validate AutoSco’s effectiveness in automated figure skating scoring.