GS-Octree: Octree-based 3D Gaussian Splatting for Robust Object-level 3D Reconstruction Under Strong Lighting

TL;DR

Experimental results show that the method, which leverages the distribution of 3D Gaussians with SDFs, reconstructs more accurate geometry, particularly in images with specular highlights caused by strong lighting, reconstructs more accurate geometry.

Abstract

The 3D Gaussian Splatting technique has significantly advanced the construction of radiance fields from multi-view images, enabling real-time rendering. While point-based rasterization effectively reduces computational demands for rendering, it often struggles to accurately reconstruct the geometry of the target object, especially under strong lighting. To address this challenge, we introduce a novel approach that combines octree-based implicit surface representations with Gaussian splatting. Our method consists of four stages. Initially, it reconstructs a signed distance field (SDF) and a radiance field through volume rendering, encoding them in a low-resolution octree. The initial SDF represents the coarse geometry of the target object. Subsequently, it introduces 3D Gaussians as additional degrees of freedom, which are guided by the SDF. In the third stage, the optimized Gaussians further improve the accuracy of the SDF, allowing it to recover finer geometric details compared to the initial SDF obtained in the first stage. Finally, it adopts the refined SDF to further optimize the 3D Gaussians via splatting, eliminating those that contribute little to visual appearance. Experimental results show that our method, which leverages the distribution of 3D Gaussians with SDFs, reconstructs more accurate geometry, particularly in images with specular highlights caused by strong lighting.

Figures (13)

• Figure 1: Our method integrates octree-based implicit surface representations with Gaussian splatting, enabling real-time rendering for novel view synthesis using fewer Gaussians. This integration allows us to robustly reconstruct high-quality geometry from input images with large areas of specular highlight due to strong lighting. The values below each figure represent the Chamfer distance ($10^{-4}$), PSNR, FPS and the number of Gaussians $N_{\mathrm{GS}}$ (in thousands), with the best results highlighted in bold.
• Figure 2: Algorithmic pipeline. Our framework consists mainly of four stages. Utilizing the octree, we progressively optimize the signed distance values $s_{ij}$ for geometry and the SH coefficients $\mathbf{a}_{ij}$ for radiance for each octree node in a coarse-to-fine manner, alternating between volume rendering and point splatting.
• Figure 3: Qualitative results from octrees spanning level 6 to 9. As the resolution of the octree increases, it provides more degrees of freedom, effectively improving the quality of the reconstructed geometry.
• Figure 4: Gaussian-guided geometric optimization. Left: Pure octree-based SDFs are prone to inaccurate geometry, and simply refining the octree does not address the issue. Right: We propose to leverage a Gaussian point cloud generated from GS to enhance reconstruction accuracy.
• Figure 5: We calculate the Hessian matrix using a numerical method based on octree grids.