PURPOSE : This study aimed to compare the performance and efficiency of sequence models — LSTM and GRU — using both 2D and 3D motion data for classifying archery shooting techniques. The goal was to evaluate how data dimensionality and model architecture influence classification accuracy, computational cost, and practical applicability in sports motion analysis. METHODS : Four models (2D LSTM, 2D GRU, 3D LSTM, and 3D GRU) were trained and tested on motion capture data collected from elite archers. The dataset was split into training and testing sets at a 7:3 ratio, with 6,304 samples for 2D data and 2,317 samples for 3D data. Performance was assessed using accuracy, loss, F1-score, and AUC, while training time was recorded to evaluate computational efficiency. Confusion matrices were analyzed to identify class-level strengths and weaknesses. RESULTS : The GRU-based models consistently outperformed LSTM-based models across both 2D and 3D datasets. The 3D GRU achieved the highest accuracy (89%) and F1-score, benefiting from depth and spatial information, while the 2D GRU reached 82% accuracy with the shortest training time (20 minutes). The performance gain from 2D to 3D was greater for GRU models (+7 percentage points) than for LSTM models (+3 percentage points). Confusion matrix analysis revealed that 3D models, particularly the 3D GRU, exhibited stronger class separation, though 2D models maintained competitive performance with lower computational demands. CONCLUSIONS : While 3D GRU offers the highest classification accuracy, its longer training time and higher computational cost may limit its suitability for real-time or resource-constrained applications. The 2D GRU model provides a practical balance between performance and efficiency, making it well-suited for rapid analysis and frequent model updates. The choice between 2D and 3D approaches should be guided by the intended application, available resources, and the required balance between accuracy and efficiency.
© Copyright 2025 Journal of Asian Society for Health & Exercise. All rights reserved.