highlight-generation-icehockey

SimulaMet-HOST/highlight-generation-icehockey

highlight-generation-icehockey is a image classification model from SimulaMet-HOST. Use it when you need a label for an image. The card lists the license as h-research.

<a href="-overview"Overview</a • <a href="-github-repository"GitHub Repository</a • <a href="-classes"Classes</a • <a href="-dataset"Dataset</a • <a href="-models"Models</a • <a href="-intended-use"Intended Use</a • <…

Downloads · 30 days

Access

Public

Updated Jul 2, 2026

Repo size

2.4 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.pth2.4 GB · 100%

At a glance

Task: Image Classification
License: h-research
Access: Public
Created: Jun 26, 2026
Updated: Jul 2, 2026
SHA: 9e652b62

Model card

From the Hugging Face model README

Ice Hockey Shot-Type Classifiers

<a href="#-overview">Overview</a> • <a href="#-github-repository">GitHub Repository</a> • <a href="#-classes">Classes</a> • <a href="#-dataset">Dataset</a> • <a href="#-models">Models</a> • <a href="#-intended-use">Intended Use</a> • <a href="#-citation">Citation</a> • <a href="#-contact">Contact</a>

Overview

This repository contains shot-type classification models trained on Swedish Hockey League broadcast footage. The models are used in an automatic ice hockey highlight generation pipeline, where structured goal events are converted into edited goal highlights.

The classifier labels broadcast frames by camera view. These labels help the pipeline identify the main goal sequence, reaction shots, staff or player close-ups, crowd views, and behind-the-goal material before the final video is assembled.

GitHub Repository

The source code for the automatic ice hockey highlight generation pipeline is available on GitHub:

https://github.com/forzasys-students/highlight-generation-icehockey

Classes

The models classify frames into seven camera-view classes.

Class	Meaning
`close_up_player_or_field_referee`	Close-up of players or referees
`close_up_side_or_staff`	Bench, staff, or side-area close-up
`main_camera_left`	Main camera, left view
`main_camera_center`	Main camera, center view
`main_camera_right`	Main camera, right view
`behind_the_goal`	View from behind or near the goal
`public_or_fans`	Crowd or fan shot

Dataset

The dataset contains 15,005 annotated frames from SHL broadcast material.

Class	Frames
`close_up_player_or_field_referee`	4,440
`close_up_side_or_staff`	2,833
`main_camera_left`	2,420
`main_camera_center`	2,177
`main_camera_right`	1,492
`behind_the_goal`	968
`public_or_fans`	675
Total	15,005

Models

We trained and evaluated several image-classification backbones, including EfficientNet, ResNet, MobileNetV3, ConvNeXt, and ViT variants.

Model	Accuracy	Macro F1
🥇 ConvNeXt-Tiny	0.983	0.982
EfficientNet-B1	0.979	0.978
EfficientNet-B3	0.978	0.976
EfficientNet-B2	0.975	0.976
EfficientNet-B0	0.975	0.973
ResNet50	0.973	0.972
ViT-B-16	0.975	0.972
MobileNetV3-Large	0.971	0.971
ResNet34	0.971	0.969
ResNet18	0.963	0.965

Intended Use

These models are intended for research on ice hockey broadcast analysis.

They may need retraining or adaptation for other leagues, arenas, camera setups, or broadcast styles.

Contact

For questions, please contact:

Name	Email
Mehdi Houshmand	[email protected]
Pål Halvorsen	[email protected]
Cise Midoglu	[email protected]

Task: Image Classification
License: h-research
Languages: en
Created: Jun 26, 2026
Updated: Jul 2, 2026