Unbounded-GS: Extending 3D Gaussian Splatting With Hybrid Representation for Unbounded Large-Scale Scene Reconstruction

Modeling large-scale scenes from multi-view images is challenging due to the trade-off dilemma between visual quality and computational cost. Existing NeRF-based methods have made advancements in neural implicit representation through volumetric ray-marching, but still struggle to deal with cubically growing sampling space in large-scale scenes. Fortunately, the rendering approach based on 3D Gaussian splatting (3DGS) has shown promising results, inspiring further exploration in the splatting setting. However, 3DGS has the limitation of inadequate Gaussian points for modeling distant backgrounds, leading to “splotchy” artifacts. To address this problem, we introduce a novel hybrid neural representation called Unbounded 3D Gaussian. For foreground area, we employs an explicit 3D Gaussian representation to efficiently model the geometry and appearance through splatting weighted Gaussians. For far-away background, we additionally introduce an implicit module comprising Multi-layer Perceptions (MLPs) to directly predict far-away background colors from positional encodings of view positions and ray directions. Furthermore, we design a seamless blending mechanism between the color predictions of the explicit splatting and implicit branches to reconstruct holistic scenes. Extensive experiments demonstrate that our proposed Unbounded-GS inherits the advantages of both faster convergence and high-fidelity rendering quality.

I. Introduction

Modeling 3D scenes from multi-view images drives advancements in robotic perception and navigation, enabling tasks such as inspection operations, path planning and obstacle avoidance [1], [2], [3], [4]. These scenarios demand highly accurate and high-resolution 3D representations to ensure effective performance [2]. Recent methods, like Neural Radiance Fields (NeRF) [5], have excelled in object-centered scene reconstruction and novel view synthesis, making them valuable for generating photorealistic views. Extensions of NeRF to large-scale scenes [4], [6] are particularly relevant for modeling expansive environments necessary for robotic applications. However, the significant computational overhead of these methods, such as the extended rendering times required by Mip-NeRF 360 [6], limits their practical use in robotics, where real-time processing is essential. To mitigate these demands, auxiliary explicit voxel grids have been employed to encode local features more efficiently [7], [8], [9]. While these approaches reduce computational costs, they often compromise visual quality, which is critical for precise robotic tasks.