DynamicGS: Spatiotemporal Motion Distillation for Dynamic Scene Reconstruction with Gaussian Splatting

Dynamic scene reconstruction is essential for industrial vision and robotic perception, where accurate and efficient modeling of time-varying geometry is crucial for applications such as automated inspection, robotic manipulation, and autonomous navigation. 3D Gaussian Splatting (3DGS), with its explicit Gaussian primitives and efficient rasterization, offers an effective solution for high-quality, real-time scene reconstruction. However, existing 3DGS-based methods often rely on dense multi-view observations, making them vulnerable to geometric and temporal instability when only sparse views or monocular videos are available. In such settings, insufficient observations severely constrain Gaussian primitives, leading to inaccurate motion estimation, structural deformation, and temporal inconsistency. To address these challenges, we propose DynamicGS, a spatiotemporal motion distillation framework for dynamic scene reconstruction with Gaussian Splatting. The proposed DynamicGS explicitly models and regularizes Gaussian motion to improve the geometric stability and temporal consistency of reconstructed scenes. First, we introduce a canonical space and deformation field to represent the spatiotemporal evolution of dynamic scenes. Second, we design an explicit motion-aware representation to embed dynamic motion features into Gaussian primitives, enabling more effective modeling of time-varying geometry. Third, we develop a spatiotemporal motion distillation mechanism to constrain and refine Gaussian motion from a global perspective, reducing motion ambiguity and improving temporal coherence. Extensive experiments on D-NeRF and HyperNeRF datasets show that our DynamicGS outperforms other state-of-the-art methods and effectively improves both reconstruction quality and rendering efficiency for dynamic scenes.

Authors

Institutions

Publication Details

Journal
Computers
Published
2026-11-19
DOI
https://doi.org/10.3390/computers15110854
Primary Topic
3D Shape Modeling and Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

DynamicGS: Spatiotemporal Motion Distillation for Dynamic Scene Reconstruction with Gaussian Splatting

Mingwei Cao, Haifeng Zhao, Ning Li, Zhihao Liu
Computers
3D Shape Modeling and Analysis
article

DynamicGS: Spatiotemporal Motion Distillation for Dynamic Scene Reconstruction with Gaussian Splatting

Mingwei Cao, Haifeng Zhao, Ning Li, Zhihao Liu
article en

Abstract

Dynamic scene reconstruction is essential for industrial vision and robotic perception, where accurate and efficient modeling of time-varying geometry is crucial for applications such as automated inspection, robotic manipulation, and autonomous navigation. 3D Gaussian Splatting (3DGS), with its explicit Gaussian primitives and efficient rasterization, offers an effective solution for high-quality, real-time scene reconstruction. However, existing 3DGS-based methods often rely on dense multi-view observations, making them vulnerable to geometric and temporal instability when only sparse views or monocular videos are available. In such settings, insufficient observations severely constrain Gaussian primitives, leading to inaccurate motion estimation, structural deformation, and temporal inconsistency. To address these challenges, we propose DynamicGS, a spatiotemporal motion distillation framework for dynamic scene reconstruction with Gaussian Splatting. The proposed DynamicGS explicitly models and regularizes Gaussian motion to improve the geometric stability and temporal consistency of reconstructed scenes. First, we introduce a canonical space and deformation field to represent the spatiotemporal evolution of dynamic scenes. Second, we design an explicit motion-aware representation to embed dynamic motion features into Gaussian primitives, enabling more effective modeling of time-varying geometry. Third, we develop a spatiotemporal motion distillation mechanism to constrain and refine Gaussian motion from a global perspective, reducing motion ambiguity and improving temporal coherence. Extensive experiments on D-NeRF and HyperNeRF datasets show that our DynamicGS outperforms other state-of-the-art methods and effectively improves both reconstruction quality and rendering efficiency for dynamic scenes.

ComputersVol. 15(11)
Anhui University (CN)
3D Shape Modeling and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon