BenderNet and RingerNet: Highly Efficient Line Segmentation Deep Neural Network Architectures for Ice Rink Localization

Abstract

A critical step for computer vision-driven hockey ice rink localization from broadcast video is the automatic segmentation of lines on the rink. While the leveraging of segmentation methods for sports field localization has been previously explored, the design of deep neural networks for segmenting ice rink lines has not been well studied. Furthermore, the exploration of efficient architecture designs is very important given the operational requirements of real-time sports analytics. Motivated by this, BenderNet and RingerNet, two highly efficient deep neural network architectures, have been designed specifically for ice rink line segmentation. Experiments on a dataset of annotated NHL broadcast video demonstrate high accuracy while maintaining high model efficiency, thus making the proposed methods well-suited for real-time ice hockey rink localization.