Sourse: School of Intelligent Systems Science and Engineering / Industry School of Artificial Intelligence 

Publisher: Jinan University Media Center

Text: School of Intelligent Systems Science and Engineering / Industry School of Artificial Intelligence 

Editor: Su Qianyi

Images: School of Intelligent Systems Science and Engineering / Industry School of Artificial Intelligence 

Video: School of Intelligent Systems Science and Engineering / Industry School of Artificial Intelligence 

A research team led by Professor Yang Guanghua from the School of Intelligent Systems Science and Engineering/Industry School of Artificial Intelligence at Jinan University has made significant progress in city-scale real-scene 3D reconstruction. Their paper, titled “Robust and Efficient 3D Gaussian Splatting for Urban Scene Reconstruction,” has been accepted by ICCV 2025, a top-tier international conference on computer vision. The paper lists Zhen-sheng Yuan, a 2022 master’s student, as first author, Professor Yang Guanghua and Dr. Di Wang as corresponding authors, and Dr. Hao-zhi Huang and 2022 master’s student Zhen Xiong as key co-authors.

The study addresses multiple challenges in city-scale 3D reconstruction using 3D Gaussian Splatting (3DGS), such as long reconstruction times, high memory consumption, low rendering frame rates, strong appearance variation under complex lighting, interference from dynamic objects, and noticeable artifacts. Through multi-faceted technical innovations, the team has achieved state-of-the-art reconstruction quality alongside notable gains in efficiency.

Key innovations introduced in the work include:

•An intelligent visibility-based data partitioning strategy, which uses a novel visibility assessment algorithm to accurately select image data that contributes to reconstruction, greatly reducing redundancy during training and improving reconstruction efficiency.

•A partition-priority dynamic resource allocation mechanism that optimizes memory usage and speeds up reconstruction, alongside a controllable Level of Detail (LOD) generation strategy enabling bottom-up multi-scale modeling and real-time rendering of city-scale scenes.

•A fine-grained appearance modeling approach that uses Gaussian primitives as basic units to represent scene appearance changes, improving robustness to variations in lighting and season.

•Multi-dimensional regularization techniques combining scale and depth cues, integrating advantages of Mip-Splatting and AbsGS to enhance detail recovery, along with object detection and semantic segmentation models to filter transient objects and further improve fidelity and robustness.

Experimental results across multiple scenes show that the proposed approach produces fewer artifacts and recovers finer details, achieving top-tier reconstruction quality. At the same time, the method maintains reasonable resource demands, enabling real-time rendering under 24GB or lower GPU memory, demonstrating strong potential for practical city-scale applications.

(LOD-related diagram annotations: Quality Metric, Efficiency Metric; Previous Method, Our Method; Previous Method, Our Method; Fewer artifacts, stronger detail recovery; (a) Rubble, (b) Campus, (c) BigCity)

Building on these advancements, the team has developed a “Digital Base for Low-Altitude Economy” platform. This system integrates key technologies, intelligent algorithms, and application scenarios into a unified architecture, serving as a digital infrastructure for future low-altitude economies. It offers three core capabilities: high-precision 3D city reconstruction, comprehensive and refined intelligent management, and support for multi-scenario low-altitude applications. The platform is already being piloted in areas such as smart tourism, UAV inspection, and low-altitude logistics.

(Jinan University Low-altitude Economy Digital Base)

ICCV is one of the three premier conferences in computer vision and is recognized as a Rank A conference by the China Computer Federation (CCF). ICCV 2025 will take place from October 19–23, 2025, in Hawaii, USA.