---
license: h-research
language:
  - en
tags:
  - computer-vision
  - image-classification
  - sports
  - ice-hockey
  - broadcast-analysis
  - highlight-generation
pipeline_tag: image-classification
---

# Ice Hockey Shot-Type Classifiers

| | | | | |
|:--:|:--:|:--:|:--:|:--:|
| ![Task](https://img.shields.io/badge/task-image%20classification-blue) | ![Sport](https://img.shields.io/badge/sport-ice%20hockey-lightgrey) | ![Frames](https://img.shields.io/badge/dataset-15,005%20frames-green) | ![Classes](https://img.shields.io/badge/classes-7-orange) | ![License](https://img.shields.io/badge/license-h--research-purple) |

<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](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 | mehdi@forzasys.com |
| Pål Halvorsen | paalh@simula.no |
| Cise Midoglu | cise@forzasys.com |

---