Occluded and Deformed Jersey Numbers Recognition by Hourglass Networks with Deformable Convolutional Networks

Abstract

We address the problem of basketball jersey number recognition, where the jersey numbers may be partially occluded, deformed, or bent. These challenges make it difficult to accurately recognize the jersey numbers, especially in real-world scenarios with varying player postures, angles, lighting conditions, jersey colors, and patterns. In this paper, we propose a novel method for basketball jersey number recognition, name deformable hourglass network (DHN), by integrating deformable convolutional network v3 (DCNv3) into the hourglass architecture, which is inspired by the convolutional character networks (CharNet) model. We also provide a new dataset, which contains various occluded jersey numbers in real basketball contest scenarios. We show that the DHN can achieve better recognition accuracy and robustness of the deformed basketball jersey number in challenging real-world scenarios.