面向挑战性重建场景的相机参数与3D高斯联合优化

Joint Optimization of Camera Parameters and 3D Gaussians for Challenging Reconstruction Scenarios

摘要: 3D高斯泼溅(3D Gaussian Splatting, 3DGS)凭借其高效的实时渲染能力与快速收敛特性，已成为新视角合成领域的重要技术路线。然而，其重建质量高度依赖于精确的相机内参与外参标定，这些参数传统上依赖运动恢复结构(SfM)流程(如COLMAP)预先估计。值得注意的是，在低纹理、重复纹理或视角受限等挑战性场景中，SfM估计往往产生误差，限制了重建精度。为此，本文提出一种在训练阶段同步精细化相机参数与3D高斯表示的联合优化框架。具体而言，我们通过偏移参数化策略增强优化稳定性，采用对数尺度焦距表示改善数值条件，并引入余弦衰减自适应正则化项来优化相机参数，提升建模质量。

Abstract: 3D Gaussian Splatting (3DGS) has emerged as a prominent technique in novel view synthesis, distinguished by its efficient real-time rendering capabilities and rapid convergence properties. However, its reconstruction quality heavily relies on accurate estimation of camera intrinsics and extrinsics, which are traditionally obtained through Structure-from-Motion (SfM) pipelines such as COLMAP. Notably, in challenging scenarios characterized by low-texture regions, repetitive patterns, or limited viewpoint distributions, SfM estimations often suffer from inaccuracies, thereby compromising reconstruction fidelity. To address this limitation, we propose a joint optimization framework that simultaneously refines camera parameters and 3D Gaussian representations during the training stage. Specifically, we employ offset parameterization to stabilize optimization, adopt logarithmic focal length representation for improved numerical conditioning, and introduce a cosine annealing adaptive regularization strategy to refine camera parameters and enhance modeling quality.