A Transformer-based Object Relationship Finder for Object Status Analysis

Abstract

Basketball analysis systems are essential tools in modern basketball, where identifying the ball handler is one of the most critical tasks. The reason for this challenge comes from the overlapping of players in basketball, which makes it easy for the analysis system to misjudge the ball handler. We found that it is easy to misjudge ball handler using traditional algorithms, such as calculating the degree of intersection over the union or calculating the coordinate distance between the player and the ball. In this paper, we propose a transformer-based object relationship finder to classify the relationship between players and the ball, which uses features of different objects, such as the use of coordinate information and skeleton information as inputs, to learn the relationship between players and the ball through self-attention. Experimental results show that our method achieves an accuracy of ball handler up to 91.2% based on a smaller dataset, surpassing the 83.9% accuracy of traditional algorithms and the 77.8% accuracy of Resnet-based convolutional neural networks.