MobileGaussian: Efficient 3D Gaussian-based Open Vocabulary Scene Understanding on Mobile Devices
Year: 2025
Type: conference-paper
Abstract: We present MobileGaussian, a lightweight and efficient framework for open-vocabulary 3D scene understanding, designed to meet the constraints of mobile and resource-limited platforms. Building upon 3D Gaussian Splatting, our approach introduces a semantic-aware point pruning strategy to eliminate redundant data while preserving critical structural and semantic information. Additionally, we propose a semantic-weighted voting mechanism to enhance openvocabulary segmentation across diverse scenes. Extensive experiments on the Mip-NeRF360 dataset demonstrate that our method achieves significant memory savings with minimal performance degradation, delivering competitive results in terms of visual fidelity and semantic accuracy. MobileGaussian offers a practical solution for deploying advanced 3D perception in realworld applications such as AR, robotics, and mobile computing.
Authors: Xixiao Liu, Dianxi Shi, Shaowu Yang
Institution: National University of Defense Technology
Language: English
Subfield: Computational Mechanics
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4412617917