SREGS: Sparse-view Gaussian radiance fields with geometric regularization and region exploration.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Novel-View Synthesis

Background:

• Few-shot novel-view synthesis using 3D Gaussian Splatting (3DGS) shows promise.
• Existing methods struggle with geometric supervision and robustness from sparse viewpoints.

Purpose of the Study:

• To introduce SREGS, a framework for few-shot reconstruction overcoming limitations of existing methods.
• To enhance geometric consistency and accuracy in novel-view synthesis.

Main Methods:

• Explicitly consistent geometry initialization using 2D Gaussians.
• Region-adaptive densification and opacity-aware noise for improved exploration.
• Multi-scale depth-guided optimization using monocular depth priors.

Main Results:

• SREGS achieves higher synthesis quality compared to existing methods.
• Demonstrates robust performance across various datasets (LLFF, MipNeRF360, Blender).
• Achieves improved results with lower computational cost.

Conclusions:

• SREGS offers a robust and efficient solution for few-shot novel-view synthesis.
• The framework effectively addresses challenges in geometric reconstruction from sparse viewpoints.
• The proposed methods enhance both consistency and accuracy in 3D scene reconstruction.