Rink-Agnostic Hockey Rink Registration

TL;DR

This work tackles the problem of rink registration across diverse hockey rink types, aiming to generalize beyond NHL templates. It introduces a three-module pipeline that first segments rink features, then estimates a homography to warp an overhead rink template, and finally refines the warp, with training that leverages domain adaptation and synthetic data to bridge the NHL-non-NHL gap. Key contributions include domain-adaptive segmentation with augmented realism, a synthetic-data-trained homography estimator, and an iterative refinement stage that achieves competitive performance on NHL data while substantially improving non-NHL generalization. The approach offers a practical path to rink-agnostic warping for player tracking and analytics without requiring extensive labeled data for every rink type, enhancing the scalability of broadcast-video hockey analytics.

Abstract

Hockey rink registration is a useful tool for aiding and automating sports analysis. When combined with player tracking, it can provide location information of players on the rink by estimating a homography matrix that can warp broadcast video frames onto an overhead template of the rink, or vice versa. However, most existing techniques require accurate ground truth information, which can take many hours to annotate, and only work on the trained rink types. In this paper, we propose a generalized rink registration pipeline that, once trained, can be applied to both seen and unseen rink types with only an overhead rink template and the video frame as inputs. Our pipeline uses domain adaptation techniques, semi-supervised learning, and synthetic data during training to achieve this ability and overcome the lack of non-NHL training data. The proposed method is evaluated on both NHL (source) and non-NHL (target) rink data and the results demonstrate that our approach can generalize to non-NHL rinks, while maintaining competitive performance on NHL rinks.

Figures (15)

• Figure 1: Example of warping a video frame onto the overhead rink template (and vice versa) using homography.
• Figure 2: Examples of different rinks. On top of the differences between rink shape and feature positioning, there are also differences in color, advertising frequency, and how face-off circles were filled. Face-off circle differences are highlighted using dashed boxes
• Figure 3: Pipeline of the process during test time, showing the 3 major components. The inputs to the pipeline are the video frame fed to the segmentation model and the overhead template fed to the initial estimator. The iteration of the refinement model has been omitted for clarity.
• Figure 4: The line and segmentation overhead templates used for NHL (left) and Olympic (right) rinks. Note that in reality, both rinks are the same lengthwise, and the Olympic rinks are wider than the NHL rinks. They were both scaled to fit the same template space for this analysis.
• Figure 5: Examples of logo augmentation, which sometimes added text, rectangles, and circle fillings in order to augment the existing dataset further.