Introducing Unbiased Depth into 2D Gaussian Splatting for High-accuracy Surface Reconstruction (2503.06587v2)

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 found 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 rectified 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.

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

The paper "Introducing Unbiased Depth into 2D Gaussian Splatting for High-accuracy Surface Reconstruction" delineates an advanced methodology that builds over existing frameworks of Gaussian splatting, specifically focusing on refining the geometric accuracy of 3D reconstructions from 2D images. The key contributions of the research are centered around addressing critical challenges in the accurate representation of glossy surfaces, which are often marred by discontinuities and biased depth estimation.

Technical Innovations

The paper begins by introducing 2D Gaussian Splatting (2DGS), enhancing it with unbiased depth estimation techniques that significantly improve the surface reconstruction quality. Traditionally, 2DGS, while beneficial for approximating thin surfaces, struggles with reflective surfaces due to discontinuity caused by specular reflections, resulting in visible holes and pits. The paper identifies these issues and introduces novel components to the splatting framework:

1. Depth Convergence Loss: This loss replaces the original depth distortion loss. It imposes a robust constraint on depth continuity across glossy regions, mitigating the creation of concave Gaussian splats and fostering an "edge-growing" effect, thus ensuring more accurate adherence to the actual surface.
2. Depth Correction Criterion: The authors propose a new criterion for determining the actual surface depth. This method considers both the number of intersecting Gaussians and the accumulated opacity, facilitating enhancement in surface representation accuracy, especially in high-specularity regions.

Numerical Results and Evaluation

The evaluations, conducted on diverse datasets such as DTU, Tanks and Temples, and Mip-NeRF360, showcase substantial improvements in geometric accuracy and novel view synthesis quality. The proposed method outperforms existing Gaussian splatting techniques, such as 3DGS, SuGaR, and the original 2DGS, on crucial metrics like Chamfer distance and PSNR, particularly in scenes with glossy surfaces where high accuracy is indispensable. The experiments explicitly demonstrate the enhanced ability of the method to rectify major artifacts associated with specular highlights.

Implications and Future Work

Both theoretical and practical implications of the paper are noteworthy. From a theoretical standpoint, the research introduces a more comprehensive understanding of Gaussian splat behavior in complex lighting conditions, pushing the boundaries of surface reconstruction methodologies. Practically, the improved geometric fidelity opens avenues for applications requiring detailed 3D models, such as digital content creation, augmented reality, and industrial design.

Looking forward, the paper sets the stage for further exploration into adaptive Gaussian splatting techniques that could efficiently handle local dense and tiny parts in complex scenes, addressing the limitations identified during evaluations on datasets like Mip-NeRF360. Additionally, integrating more sophisticated neural implicit models with unbiased depth estimation could herald even more precise surface reconstructions.

In conclusion, this paper provides a pivotal advancement in the domain of 3D surface reconstruction using Gaussian splatting, furnishing significant improvements in handling the intricacies of glossy surfaces and setting a new benchmark for future research in this direction.