AGES：各向异性高斯平滑增强的三维几何一致性重建方法

AGES: Anisotropic Gaussian Enhancement with Smoothness for Geometric-Consistent 3D Reconstruction Method

摘要: 在稀疏视角和低纹理区域条件下，从多视角图像实现高保真三维重建仍然是一个具有挑战性的问题，传统方法在此类场景中往往表现不稳定。尽管三维高斯溅射(3D Gaussian Splatting, 3DGS)能够实现实时渲染，但其几何优化与各向异性外观建模相互解耦，容易在无纹理或高反射区域产生伪影。针对上述问题，本文提出一种各向异性高斯平滑增强方法(Anisotropic Gaussian Enhancement with Smoothness, AGES)，构建了一个基于概率路由的联合优化框架，在单一训练流程中联合优化场景几何结构与视角相关外观建模。该方法引入两个关键组成部分：(1) 自适应几何–外观路由模块(Adaptive Geometry-Appearance Routing, AGAR)，基于学习得到的逐高斯不确定性度量，动态地将高斯基元分配至跨视角几何细化分支或各向异性反射建模分支；(2) 深度平滑正则化项(Depth Smoothness Regularization, DSR)，通过约束渲染深度与几何优化深度之间的局部梯度一致性，在抑制噪声的同时有效保持结构边缘。大量在Waymo和YouTube数据集上的实验结果表明，所提出的AGES方法在全分辨率及降采样设置下均显著优于现有先进方法，在复杂真实场景中实现了更高的几何一致性和视觉保真度。

Abstract: High-fidelity 3D reconstruction from multi-view images remains a significant challenge, particularly under sparse viewpoints and in low-texture regions where conventional methods are often unreliable. While 3D Gaussian Splatting (3DGS) enables real-time rendering, the decoupling of its geometric optimization and anisotropic appearance modeling frequently leads to artifacts in textureless or specular areas. To address these limitations, we propose Anisotropic Gaussian Enhancement with Smoothness (AGES), which formulates a probability-guided joint optimization framework to jointly optimize scene geometry and view-dependent appearance within a single training pipeline. Our approach introduces two key components: (1) an Adaptive Geometry-Appearance Routing (AGAR) module, which dynamically routes each Gaussian to either a cross-view geometric refinement branch or an anisotropic reflectance modeling branch based on a learned per-primitive uncertainty measure; and (2) a Depth Smoothness Regularization (DSR) loss, which enforces local gradient consistency between the rendered and geometrically refined depth maps to preserve structural edges while suppressing noise. Extensive experiments on the Waymo and YouTube datasets demonstrate that AGES significantly outperforms state-of-the-art methods at both full and downsampled resolutions, achieving superior geometric consistency and visual fidelity in challenging real-world scenarios.