Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

4D Sports Scene Reconstruction: From Motion Capture System to Free-Viewpoint Video using Gaussian Splatting

Jiao, Keming LU (2026) MAMM15 20261
Ergonomics and Aerosol Technology
Department of Design Sciences
Abstract
Key actions, tactical cooperation, and spatial relationships in sports activities are usually recorded through 2D videos, which limits users’ ability to understand the game in three-dimensional space. Although professional sports events can provide Free-Viewpoint Video using large-scale camera arrays, such systems are difficult to deploy in low-cost scenarios such as community games and personal training. To address this problem, this thesis presents a 4D sports scene reconstruction and replay pipeline based on motion capture data and 3D Gaussian Splatting representations. The system takes sparse multi-view videos captured by a small number of fixed cameras as input, together with camera calibration, human segmentation, and estimated human... (More)
Key actions, tactical cooperation, and spatial relationships in sports activities are usually recorded through 2D videos, which limits users’ ability to understand the game in three-dimensional space. Although professional sports events can provide Free-Viewpoint Video using large-scale camera arrays, such systems are difficult to deploy in low-cost scenarios such as community games and personal training. To address this problem, this thesis presents a 4D sports scene reconstruction and replay pipeline based on motion capture data and 3D Gaussian Splatting representations. The system takes sparse multi-view videos captured by a small number of fixed cameras as input, together with camera calibration, human segmentation, and estimated human pose parameters. It reconstructs dynamic human Gaussian avatars and the static environment separately, and integrates them in a unified interactive viewer for scene composition and rendering. To improve sparse-view multi-person reconstruction, the method incorporates pose correction, non-uniform Gaussian initialization, two-stage training, and a perceptually oriented loss-weight setting. The avatar reconstruction component is evaluated on the GalaBasketball and Hi4D datasets using objective metrics, subjective assessment, and ablation studies. Environment reconstruction and full scene replay are demonstrated separately using BASKET-Multiview data with more background views. The results validate the effectiveness of sparse-view avatar reconstruction and demonstrate the feasibility of the complete replay pipeline, while a fully low-camera end-to-end system remains a direction for future deployment. The thesis also discusses limitations in detail recovery, training efficiency, and the use of generative teacher models. (Less)
Please use this url to cite or link to this publication:
author
Jiao, Keming LU
supervisor
organization
course
MAMM15 20261
year
type
H2 - Master's Degree (Two Years)
subject
language
English
id
9231164
date added to LUP
2026-06-04 13:17:33
date last changed
2026-06-09 13:13:45
@misc{9231164,
  abstract     = {{Key actions, tactical cooperation, and spatial relationships in sports activities are usually recorded through 2D videos, which limits users’ ability to understand the game in three-dimensional space. Although professional sports events can provide Free-Viewpoint Video using large-scale camera arrays, such systems are difficult to deploy in low-cost scenarios such as community games and personal training. To address this problem, this thesis presents a 4D sports scene reconstruction and replay pipeline based on motion capture data and 3D Gaussian Splatting representations. The system takes sparse multi-view videos captured by a small number of fixed cameras as input, together with camera calibration, human segmentation, and estimated human pose parameters. It reconstructs dynamic human Gaussian avatars and the static environment separately, and integrates them in a unified interactive viewer for scene composition and rendering. To improve sparse-view multi-person reconstruction, the method incorporates pose correction, non-uniform Gaussian initialization, two-stage training, and a perceptually oriented loss-weight setting. The avatar reconstruction component is evaluated on the GalaBasketball and Hi4D datasets using objective metrics, subjective assessment, and ablation studies. Environment reconstruction and full scene replay are demonstrated separately using BASKET-Multiview data with more background views. The results validate the effectiveness of sparse-view avatar reconstruction and demonstrate the feasibility of the complete replay pipeline, while a fully low-camera end-to-end system remains a direction for future deployment. The thesis also discusses limitations in detail recovery, training efficiency, and the use of generative teacher models.}},
  author       = {{Jiao, Keming}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{4D Sports Scene Reconstruction: From Motion Capture System to Free-Viewpoint Video using Gaussian Splatting}},
  year         = {{2026}},
}