Robust and High-Fidelity 3D Gaussian Splatting: Fusing Pose Priors and Geometry Constraints for Texture-Deficient Outdoor Scenes

Meijun Guo, Yongliang Shi, Caiyun Liu, Yixiao Feng, Ming Ma, Tinghai Yan, Weining Lu, Bin Liang

Introduction

3D Gaussian Splatting (3DGS) has emerged as a powerful technique for real-time neural rendering. However, its application in large-scale outdoor environments faces significant hurdles. Traditional 3DGS pipelines rely on Structure-from-Motion (SfM) tools like COLMAP for camera pose estimation. In environments with repetitive or weak textures—such as long stretches of asphalt or uniform brick walls—SfM often fails to find reliable matches, leading to broken trajectories or metric drift. Furthermore, without explicit geometric constraints, the resulting "splats" can become poorly oriented or "needle-like," compromising the structural integrity required for robotics applications like collision detection.

Methodology

To address these challenges, we propose a robust 3DGS pipeline that integrates LiDAR-Inertial Odometry (LIO) as a pose prior and introduces novel geometric constraints to ensure physical consistency.

Overview of the proposed 3DGS pipeline, integrating LIO pose priors and geometric constraints.

Scale-Aware Visual Pose Optimization

Rather than initiating SfM from scratch, our method utilizes FAST-LIO2 to provide a metric-aware initial guess. These LIO poses are incorporated into the triangulation process. We formulate a bundle adjustment optimization that includes both initial pose constraints and relative constraints between adjacent frames, effectively stabilizing the trajectory and maintaining metric scale—a critical requirement for robotics that standard visual SfM often fails to meet.

Geometric Texture Consistency

To ensure that the Gaussians accurately represent physical surfaces, we introduce two new loss functions:

  1. Normal Constraints ($L_{normal}$): We leverage a pre-trained Omnidata model to predict surface normals from images. The Gaussians' minor axes are forced to align with these predicted normals, supplemented by a smoothness term to ensure spatial consistency.
  2. Shape Constraints ($L_{shape}$): We apply an "effective rank" regularization based on entropy. This mathematical constraint encourages Gaussians to adopt "disk-like" (flat) shapes rather than degenerate "needle-like" or "spherical" forms, better approximating real-world geometry.

Experimental Evaluation

The proposed method was evaluated on the public HKU-Campus dataset and a new proprietary QY Dataset, the latter specifically designed to test performance in texture-deficient campus environments where traditional COLMAP typically fails.

Key Results

  • Robustness: On the texture-deficient QY dataset, where standard COLMAP failed entirely, our method successfully reconstructed the scene with high metric accuracy.
  • Visual Fidelity: Our approach outperformed baselines such as Splatfacto and 2DGS in terms of PSNR and SSIM. On the HKU dataset, we achieved a PSNR of 27.238, a significant improvement over the original 3DGS (~25.8).
  • Efficiency: By providing high-quality priors, the system achieves stable pose solutions in approximately one-third of the time required by traditional methods, proving that better priors improve both quality and efficiency.

Conclusion

This work demonstrates that fusing LIO pose priors with geometric constraints significantly enhances the robustness and fidelity of 3D Gaussian Splatting in challenging outdoor scenes. By maintaining metric scale and enforcing surface-like geometry, the proposed pipeline makes 3DGS a viable tool for the Real2Sim2Real paradigm in robotics. Future work will focus on reducing the dependency on LiDAR and moving toward real-time, on-the-fly reconstruction.

Summary updated

BibTeX
@inproceedings{guo2025robust,
  author = {Meijun Guo and Yongliang Shi and Caiyun Liu and Yixiao Feng and Ming Ma and Tinghai Yan and Weining Lu and Bin Liang},
  title = {{Robust and High-Fidelity 3D Gaussian Splatting: Fusing Pose Priors and Geometry Constraints for Texture-Deficient Outdoor Scenes}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11246464},
}
Figure