LoopSparseGS: Loop-Based Sparse-View Friendly Gaussian Splatting.

Area of Science:

• Computer Vision
• Computer Graphics

Background:

• Photorealistic novel view synthesis (NVS) using 3D Gaussian splatting (3DGS) suffers from performance degradation with sparse input views.
• This degradation stems from limited initial points, insufficient geometric supervision, and poor regularization of Gaussian ellipsoids.

Purpose of the Study:

• To introduce LoopSparseGS, a novel framework designed to improve 3DGS performance specifically for sparse-input NVS.
• To address the limitations of existing 3DGS methods in scenarios with limited visual data.

Main Methods:

• A loop-based Progressive Gaussian Initialization (PGI) strategy iteratively densifies point clouds using rendered pseudo-images.
• Depth-alignment Regularization (DAR) leverages Structure from Motion (SfM) and monocular depth for precise geometric supervision.
• A Sparse-friendly Sampling (SFS) strategy mitigates errors caused by oversized Gaussian ellipsoids.

Main Results:

• LoopSparseGS demonstrates superior performance compared to state-of-the-art methods in sparse-input NVS.
• The framework achieves high-quality synthesis across diverse scenes, including indoor, outdoor, and object-level datasets with varying resolutions.

Conclusions:

• LoopSparseGS effectively overcomes the challenges of sparse-input NVS by introducing innovative initialization, regularization, and sampling techniques.
• The proposed method offers a robust solution for generating high-fidelity novel views even with limited input imagery.