Reproduce “Initialize to Generalize: A Stronger Initialization Pipeline for Sparse-View 3DGS” on OpenResearch

About this paper

Feng Zhou, Wenkai Guo, Pu Cao, Zhicheng Zhang, Jianqin Yin

Researchers at Beijing University of Posts and Telecommunications developed 'Initialize to Generalize' (ItG-GS), a robust initialization pipeline for sparse-view 3D Gaussian Splatting. This method improves novel view synthesis quality by generating a denser and more reliable initial point cloud, achieving state-of-the-art performance on benchmarks like Mip-NeRF360 with a PSNR of 20.07 when combined with DropGS.

Key insights

• The quality of the initial 3D point cloud is the decisive factor in determining the achievable performance ceiling for sparse-view 3D Gaussian Splatting.
• Training-time regularization methods, while offering modest improvements, do not fundamentally shift the performance ceiling set by the initial point cloud.
• A hybrid initialization approach that combines relaxed SfM for robust geometric seeds, 3DGS self-supervision for dense photometric lifting, and rigorous point cloud regularization is more effective than relying solely on either traditional SfM or in-training regularization.

Summary by alphaXiv.