SoccerSynth-Detection: A Synthetic Dataset for Soccer Player Detection

Haobin Qin, Calvin Yeung, Rikuhei Umemoto, Keisuke Fujii · 2025

Read SoccerSynth-Detection: A Synthetic Dataset for Soccer Player Detection by Haobin Qin, Calvin Yeung, Rikuhei Umemoto, Keisuke Fujii (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-09281.txt.

Introduction

Haobin Qin and colleagues built a synthetic dataset in 2025 for detecting soccer players in video. Player detection underpins event spotting and tactical analysis, but real match footage is scarce because of copyright limits, crowded scenes, and constant blocking of one player by another. The few existing real datasets lack variety, so detection models adapt poorly to new footage.

Their dataset, SoccerSynth-Detection, renders artificial players under randomized lighting and textures and adds simulated camera motion blur. They tested it with a standard detection model against two real datasets. Models transferred from the synthetic data matched real-data performance overall and did clearly better on blurred images. Used as pre-training before fine-tuning, the synthetic data also improved overall detection results.

Catalog record

CREATORHaobin Qin, Calvin Yeung, Rikuhei Umemoto, Keisuke Fujii
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-09281

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SoccerSynth-Detection: A Synthetic Dataset for Soccer Player Detection Haobin Qin, Calvin Yeung, Rikuhei Umemoto, Keisuke Fujii Abstract In soccer video analysis, player detection is essential for identifying key events and reconstructing tactical positions. The presence of numerous players and frequent occlusions, combined with copyright restrictions, severely restricts the availability of datasets, leaving limited options such as SoccerNet-Tracking and SportsMOT. These datasets suffer from a lack of diversity, which hinders algorithms from adapting effectively to varied soccer video contexts. To address these challenges, we developed SoccerSynth-Detection, the first synthetic dataset designed for the detection of synthetic soccer players. It includes a broad range of random lighting and textures, as well as simulated camera motion blur. We validated its efficacy using the object detection model (Yolov8n) against real-world datasets (SoccerNet-Tracking and SportsMoT). In transfer tests, it matched the performance of real datasets and significantly outperformed them in images with motion blur; in pre-training tests, it demonstrated its efficacy as a pre-training dataset, significantly enhancing the algorithm’s overall performance. Our work demonstrates the potential of synthetic datasets to replace real datasets for algorithm training in the field of soccer video analysis. keywords synthetic datasets, object detection, soccer 1 Introduction Player detection serves as a foundational task in sports video processing, directly influencing subsequent tasks such as tracking Hu et al. 2024 ; Scott et al. 2022 , action recognition Vanderplaetse and Dupont 2020 ; Giancola et al. 2018 ; Giancola and Ghanem 2021 ; Wu et al. 2022 ; Yin et al. 2024 , action prediction Fang et al. 2024 , and game state reconstruction Somers et al. 2024 ; Ren et al. 2010 . Traditional detection methods such as the Viola-Jones Detector Viola and Jones 2001 , Histograms of Oriented Gradients Dalal and Triggs 2005 , and Discriminatively Trained Part-Based Models Felzenszwalb et al. 2009 rely on manually designed features. In contrast, deep learning-based algorithms like YOLO (You Only Look Once) Redmon 2016 ; Redmon and Farhadi 2017 ; Redmon 2018 ; Bochkovskiy et al. 2020 ; Jocher et al. 2021 ; Li et al. 2022 ; Wang et al. 2023 ; Varghese and Sambath 2024 enhance detection speed through a unique single-shot approach, offering real-time, robust detection and significant evolutionary improvements. Faster R-CNN Girshick 2015 also contributes to the field with its advanced capabilities. Despite significant advancements in detection algorithms, the application to sports scenarios encounters challenging due to complexities in annotating datasets, including frequent occlusions, dynamic backgrounds, lighting variations, motion-induced blurring, and copyright restrictions. Among various sports, soccer stands out with its large number of players, which exacerbates occlusions and complicates tracking. There are several datasets for the detection of team sports players, such as SoccerNet-Tracking Cioppa et al. 2022 , a tracking dataset from live soccer match recordings; SoccerTrack Scott et al. 2022 and TeamTrack Scott et al. 2024 , using fisheye cameras and drones to cover various sports; SoccerDB Jiang et al. 2020 , a target detection dataset from TV broadcasts; 3x3 basketball datasets Yin et al. 2024 ; Hu et al. 2024 ; and SportsMot Cui et al. 2023 , designed for multi-sport player tracking. However, the overall availability of datasets for the detection of soccer players remains limited. Employing simulators to generate synthetic datasets has emerged as a promising solution to address dataset scarcity, gaining significant attention in recent years for enabling large-scale, low-cost, and precise label generation across domains such as pedestrian detection Fabbri et al. 2021 ; Hattori et al. 2015 ; Kohl et al. 2020 , autonomous driving Sun et al. 2022 , and indoor scene understanding Roberts et al. 2021 ; Zheng et al. 2020 . In the field of soccer, several studies have attempted to create synthetic datasets. For instance, the SoccerNet-Depth dataset Leduc et al. 2024 employs NVIDIA Nsight Graphics to extract depth maps of players from eFootball games. The RADepthNet study Li et al. 2022 produced a soccer depth map dataset known as the FIFADataset, using the open-source tool RenderDoc. Web-based tools are utilized to generate soccer field datasets for keypoint extraction Fernandes and Kenwright 2022 . Furthermore, Kaneko et al. Kaneko et al. 2024 implemented the Google Research Football simulator Kurach et al. 2020 to train algorithms aimed at predicting pass movements in soccer. Additionally, Qin et al. Qin et al. 2022 outline a methodology that employs a soccer camera simulator with reinforcement learning models. However, despite these efforts, the lack of synthetic datasets for detecting soccer players persists because current simulators and games are not specifically designed for computer vision research and therefore lack the functionality to generate labeled datasets. In this paper, we release SoccerSynth-Detection , the first synthetic dataset for soccer player detection. We validated its efficacy using the Yolov8n model Varghese and Sambath 2024 , comparing it with real-world datasets: Soccernet-Tracking Cioppa et al. 2022 and SportsMoT Cui et al. 2023 . Experimental results demonstrate that models trained on SoccerSynth-Detection achieve superior generalization compared to those trained on real datasets, particularly in handling motion blur. Pretraining with SoccerSynth-Detection further enhances model performance, especially in low-data scenarios. We have also released the dataset generator of SoccerSynth-Detection, providing a powerful tool for research requiring extensive data. The contributions of this paper are as follows: 1. We release the SoccerSynth-Detection dataset, the first synthetic dataset specifically designed for soccer player detection. 2. We develop the simulator Qin et al. 2022 to generate soccer detection datasets by enhancing camera features for detection and introducing random lighting and textures. 3. We demonstrate that SoccerSynth-Detection can replace and complement real-world datasets for training detection models. 2 Method 2.1 Simulator Construction Due to the limitations of current football simulators (e.g., Kurach et al. 2020 ; Liu et al. 2019 ; Kitano et al. 1998 ) in supporting bounding box generation, we built upon the previous soccer stadium simulator Qin et al. 2022 to set up a new center camera, with its configurations based on the positions, angles, and focal lengths found in SoccerNet-Tracking Cioppa et al. 2022 and SportsMoT Cui et al. 2023 , as illustrated in Figure 1 . Additionally, we incorporated dynamic spectator models purchased from the Unreal Engine Marketplace. Figure 1: Soccer Stadium. From left to right: (a) Drone view (b) Corner camera view, and (c) Center camera view. To enhance the realism of the synthetic players, we integrated character models and animations sourced from the Unreal Engine Marketplace and implemented their movement logic using the Behavior Tree and AI Controller within Unreal Engine. Additionally, we simulated actions such as passing, dribbling, shooting, and receiving through the physics engine, enabling continuous and autonomous football matches within the stadium environment. The behavioral rules governing the players are as follows: • Possession Rules: Possession is assigned to the team of the last player who touched the football. • Actions in Possession: The goalkeeper passes to forwards; defenders to midfielders; midfielders to forwards. Forwards either pass back to midfielders if surrounded or dribble towards the penalty area and shoot at a randomly selected target within the goal. • Actions When Opponents Possess the Ball: The nearest player tries to regain possession. The goalkeeper acts to regain possession if the ball is within the penalty area. Other players engage in strategic off-ball

Read the complete text