Effects of variable pose input on neural network model performance for classifying basketball player activity

In this study, we evaluate the performance of an encoder-decoder transformer and a Spatial–Temporal Graph Convolutional Network (GCN) architecture for basketball action classification using skeletal pose data. We isolate events of dribbling, passing, shooting, and rebounding throughout each basketball game and organize player joint frame windows to represent each activity. Analyzing 82 basketball games and over 400,000 events, we demonstrate that both architectures achieve high classification accuracy even if the number of tracked joints is significantly decreased. Our experiments confirm that reducing the full body skeleton from 16 joints per player to as few as 2 joints (left and right wrists) maintains robust performance while lowering computational costs and data storage requirements—a crucial consideration in high frame-rate basketball scenarios. These findings support a streamlined approach to pose-based action recognition that has the potential to enhance real-time decision making and deployment in sporting environments.