HockeyAI: A Multi-Class Ice Hockey Dataset for Object Detection

The HockeyAI dataset is an open-source dataset designed specifically for advancing computer vision research in ice hockey. With approximately 2,100 high-resolution frames and detailed YOLO-format annotations, this dataset provides a rich foundation for tackling the challenges of object detection in fast-paced sports environments.

The dataset is ideal for researchers, developers, and practitioners seeking to improve object detection and tracking tasks in ice hockey or similar dynamic scenarios.

Dataset Overview

The HockeyAI dataset includes frames extracted from broadcasted Swedish Hockey League (SHL) games. Each frame is manually annotated, ensuring high-quality labels for both dynamic objects (e.g., players, puck) and static rink elements (e.g., goalposts, center ice).

Classes

The dataset includes annotations for the following seven classes:

• centerIce: Center circle on the rink
• faceoff: Faceoff dots
• goal: Goal frame
• goaltender: Goalkeeper
• player: Ice hockey players
• puck: The small, fast-moving object central to gameplay
• referee: Game officials

Key Highlights:

• Resolution: 1920×1080 pixels
• Frames: ~2,100
• Source: Broadcasted SHL videos
• Annotations: YOLO format, reviewed iteratively for accuracy
• Challenges Addressed:
  • Motion blur caused by fast camera movements
  • Small object (puck) detection
  • Crowded scenes with occlusions

Applications

The dataset supports a wide range of applications, including but not limited to:

• Player and Puck Tracking: Enabling real-time tracking for tactical analysis.
• Event Detection: Detecting goals, penalties, and faceoffs to automate highlight generation.
• Content Personalization: Dynamically reframing videos to suit different screen sizes.
• Sports Analytics: Improving strategy evaluation and fan engagement.

How to Use the Dataset

1. Download the dataset from Hugging Face.

2. The dataset is organized in the following structure:

HockeyAI
└── frames
    └── <Unique_ID>.jpg
└── annotations
    └── <Unique_ID>.txt

3. Each annotation file follows the YOLO format:

<class_id> <x_center> <y_center> <width> <height>

All coordinates are normalized to the image dimensions.

4. Use the dataset with your favorite object detection framework, such as YOLOv8 or PyTorch-based solutions.