2D Gaussian Splatting for Geometrically Accurate Radiance Fields

TL;DR

The paper introduces 2D Gaussian Splatting (2DGS), replacing 3D Gaussian primitives with oriented 2D disks to recover geometrically accurate radiance fields from multi-view images. A differentiable renderer performs perspective-correct ray-splat intersection and rasterization, enhanced by depth distortion and normal-consistency regularizers to stabilize optimization and improve surface sharpness. 2DGS delivers high-fidelity surface geometry with view-consistent normals and depth maps, while maintaining competitive appearance rendering and real-time rendering capabilities, offering a significant improvement over prior 3D Gaussian approaches. Extensive experiments on DTU, Tanks and Temples, and Mip-NeRF360 demonstrate superior geometry reconstruction, fast training, and robust mesh extraction via TSDF fusion, establishing 2DGS as a practical and efficient solution for radiance-field geometry and rendering.

Abstract

3D Gaussian Splatting (3DGS) has recently revolutionized radiance field reconstruction, achieving high quality novel view synthesis and fast rendering speed without baking. However, 3DGS fails to accurately represent surfaces due to the multi-view inconsistent nature of 3D Gaussians. We present 2D Gaussian Splatting (2DGS), a novel approach to model and reconstruct geometrically accurate radiance fields from multi-view images. Our key idea is to collapse the 3D volume into a set of 2D oriented planar Gaussian disks. Unlike 3D Gaussians, 2D Gaussians provide view-consistent geometry while modeling surfaces intrinsically. To accurately recover thin surfaces and achieve stable optimization, we introduce a perspective-correct 2D splatting process utilizing ray-splat intersection and rasterization. Additionally, we incorporate depth distortion and normal consistency terms to further enhance the quality of the reconstructions. We demonstrate that our differentiable renderer allows for noise-free and detailed geometry reconstruction while maintaining competitive appearance quality, fast training speed, and real-time rendering.

Figures (11)

• Figure 1: Comparison of 3DGS and 2DGS. 3DGS utilizes different intersection planes for value evaluation when viewing from different viewpoints, resulting in inconsistency. Our 2DGS provides multi-view consistent value evaluations.
• Figure 2: Illustration of 2D Gaussian Splatting. 2D Gaussian Splats are elliptical disks characterized by a center point $\mathbf{p}_k$, tangential vectors $\mathbf{t}_u$ and $\mathbf{t}_v$, and two scaling factors ($s_u$ and $s_v$) control the variance. Their elliptical projections are sampled through the ray-splat intersection ( Section \ref{['sec:splatting']}) and accumulated via alpha-blending in image space. 2DGS reconstructs surface attributes such as colors, depths, and normals through gradient descent.
• Figure 3: Visual comparisons (test-set view) between our method, 3DGS kerbl3Dgaussians, and SuGaR guedon2023sugar using scenes from an real-world dataset barron2022mipnerf360. Our method excels at synthesizing geometrically accurate radiance fields and surface reconstruction, outperforming 3DGS and SuGaR in capturing sharp edges and intricate details.
• Figure 4: Qualitative comparison on the DTU benchmark jensen2014large. Our 2DGS produces detailed and noise-free surfaces.
• Figure 5: Qualitative studies for the regularization effects. From left to right – input image, surface normals without normal consistency, without depth distortion, and our full model. Disabling the normal consistency loss leads to noisy surface orientations; conversely, omitting depth distortion regularization results in blurred surface normals. The complete model, employing both regularizations, successfully captures sharp and flat features.