Abstract
In order to develop solutions for automatic ice rink localization from broadcast video, a dataset with ground truth homographies is required. Hockey broadcast video does not tend to provide camera parameters for each frame, which means that they must be gathered manually. A novel tool for collecting ground truth transforms through point correspondences between each frame and an overhead view of the ice rink is presented in this paper. Through collaboration with the users of the tool, we have added features to improve accuracy and efficiency, especially in frames with few lines on the playing surface visible. A dataset of 4,262 frames has been collected, which will be used for research into automatic camera calibration techniques.
Introduction
Generating advanced hockey analytics, such as player tracking or puck speed estimation, requires precise knowledge of the camera's perspective relative to the playing surface. This process, known as ice rink localization, relies on a homography: a mathematical transformation matrix that maps 2D pixels from a broadcast video frame onto a static 3D model of the rink.
While deep learning models offer a path toward automating this localization, they require extensive "ground truth" datasets for training. Unfortunately, broadcast footage typically lacks camera metadata (pan, tilt, zoom), and existing sports datasets are often proprietary or focused on other sports. To address this gap, researchers at the University of Waterloo developed a specialized tool to facilitate the manual annotation of these essential transforms.
Methodology: The Annotation Tool
The developed Python-based tool employs a "human-in-the-loop" workflow to ensure high-quality data collection. The process begins with point correspondence, where a user identifies landmarks on a broadcast frame—such as the intersection of the goal line and the boards—and selects the corresponding location on a standard NHL rink model.
Once at least four points are selected, the tool calculates the transformation using the Direct Linear Transform (DLT) algorithm, refined by RANSAC (Random Sample Consensus) to mitigate the impact of inaccurate clicks. A critical feature of the tool is visual verification: it provides immediate feedback by warping the video frame onto the rink model. If the lines do not align perfectly, the user can adjust the points in real-time. To enhance efficiency, the team integrated keyboard shortcuts and "guideline" features to assist in annotating "open ice" areas where physical markings are sparse.
Dataset and Key Insights
The primary contribution of this work is a new, robust dataset comprising 4,262 annotated frames sourced from 15 NHL game broadcasts during the 2018-19 season. This dataset captures a wide variety of camera angles, broadcast graphics, and team-specific rink logos.
During development, the authors identified the "Blue Line Challenge": frames dominated by the blue line are particularly difficult to annotate because they lack the multiple line intersections (keypoints) necessary for a stable homography calculation. This underscores the importance of human-centric design; features like "zoom-in" capabilities and manual guidelines are vital for maintaining data quality in these challenging scenarios.
Results and Future Work
The tool has proven effective, with trained users averaging 90 seconds per frame for annotation. The resulting dataset of 4,262 frames is significantly larger than previous public efforts, such as the 395-frame World Cup Soccer dataset.
While manual annotation is subject to human error, this dataset provides a foundational resource for training future automated systems. Future work will focus on expanding the dataset, integrating automatic line segmentation to fine-tune homographies, and eventually achieving fully automated, human-free ice rink localization.
Summary updated