ICP-3DGS: SfM-free 3D Gaussian Splatting for Large-scale Unbounded Scenes

2025-06-24

AI Summary

Neural rendering methods (e.g., NeRF, 3DGS) for large-scale unbounded outdoor scenes heavily rely on Structure-from-Motion (SfM) to provide camera poses and sparse geometric priors—a major bottleneck in practical deployment. To address this, we propose the first end-to-end 3D Gaussian Splatting (3DGS) reconstruction framework that eliminates SfM entirely. Our method introduces two key innovations: (1) an ICP-guided differentiable pose estimation module that jointly optimizes camera trajectories and scene geometry-appearance; and (2) a voxelization-guided adaptive Gaussian densification strategy, significantly improving robustness under weak texture, low image overlap, and large camera motion. Evaluated on multi-scale indoor and outdoor datasets, our approach reduces pose estimation error by 37% and improves novel-view synthesis PSNR by 2.1 dB over SfM-dependent baselines, demonstrating state-of-the-art performance in SfM-free neural reconstruction.

Technology Category

• Computer Vision: 3D Computer Vision
• Intelligent Robots: State Estimation
• Machine Learning: Learning with Manifolds

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

In recent years, neural rendering methods such as NeRFs and 3D Gaussian Splatting (3DGS) have made significant progress in scene reconstruction and novel view synthesis. However, they heavily rely on preprocessed camera poses and 3D structural priors from structure-from-motion (SfM), which are challenging to obtain in outdoor scenarios. To address this challenge, we propose to incorporate Iterative Closest Point (ICP) with optimization-based refinement to achieve accurate camera pose estimation under large camera movements. Additionally, we introduce a voxel-based scene densification approach to guide the reconstruction in large-scale scenes. Experiments demonstrate that our approach ICP-3DGS outperforms existing methods in both camera pose estimation and novel view synthesis across indoor and outdoor scenes of various scales. Source code is available at https://github.com/Chenhao-Z/ICP-3DGS.

Problem

• Eliminates need for SfM in 3D Gaussian Splatting
• Improves camera pose estimation for large movements
• Enables large-scale unbounded scene reconstruction

Innovation

• ICP with optimization-based refinement for pose estimation
• Voxel-based scene densification for large-scale reconstruction
• SfM-free 3D Gaussian Splatting for unbounded scenes