Introducing Unbiased Depth into 2D Gaussian Splatting for High-accuracy Surface Reconstruction

Yixin Yang, Yang Zhou*, Hui Huang,
Visual Computing Research Center, CSSE, Shenzhen University
Pacific Graphics 2025 (Journal Track)
*Corresponding Author
Paper arXiv Code
Teaser Image

We introduce unbiased depth into 2DGS, effectively solving the holes in specular highlight regions and significantly improving the geometry quality of surface reconstruction.

Abstract

Recently, 2D Gaussian Splatting (2DGS) has demonstrated superior geometry reconstruction quality than the popular 3DGS by using 2D surfels to approximate thin surfaces. However, it falls short when dealing with glossy surfaces, resulting in visible holes in these areas. We find that the reflection discontinuity causes the issue. To fit the jump from diffuse to specular reflection at different viewing angles, depth bias is introduced in the optimized Gaussian primitives. To address that, we first replace the depth distortion loss in 2DGS with a novel depth convergence loss, which imposes a strong constraint on depth continuity. Then, we rectify the depth criterion in determining the actual surface, which fully accounts for all the intersecting Gaussians along the ray. Qualitative and quantitative evaluations across various datasets reveal that our method significantly improves reconstruction quality, with more complete and accurate surfaces than 2DGS.

Method

In regions where specular highlights appear on glossy surfaces, only a few Gaussian surfels can adhere to the actual surface in the 2DGS results. In contrast, most Gaussian surfels tend to concave inward in these areas, resulting in holes or pits in the reconstruction.

To address the discontinuity issue, we propose two key components. First, we propose a depth convergence loss which forces the Gaussians' depth to be continuous and smooth. The proposed constraint will promote an "edge-growing" effect from the specular-diffuse boundary to recover complete surface geometry. Second, we introduce depth correction, a new criterion to identify the actual surface, which considers both the number of intersecting Gaussians and the accumulated opacity along the ray.

Method Diagram

Illustration of the depressed Gaussians and the proposed depth convergence process.

Geometry Reconstruction





BibTeX

@article{yang2025unbiasedGS,
    author = {Yixin Yang and Yang Zhou and Hui Huang},
    title  = {Introducing Unbiased Depth into 2D Gaussian Splatting for High-accuracy Surface Reconstruction},
    eprint = {2503.06587},
    year   = {2025},
    archivePrefix = {arXiv},
}

Acknowledgements

This work was supported in parts by NSFC (U21B2023), ICFCRT (W2441020), GD Basic and Applied Basic Research Foundation (2023B1515120026), DEGP Innovation Team (2022KCXTD025), SZU Teaching Reform Key Program (JG2024018), and Scientific Development Funds from Shenzhen University.