MCGS: Multiview Consistency Enhancement for Sparse-View 3D Gaussian Radiance Fields.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• 3D Gaussian Splatting excels at novel view synthesis with efficient training and rendering.
• Sparse input views challenge existing methods due to poor Gaussian initialization and lack of multi-view consistency.
• Current approaches often rely on dense depth priors, neglecting inherent multi-view image consistency and limiting representation efficiency.

Purpose of the Study:

• To develop a view synthesis framework, MCGS, for photorealistic 3D scene reconstruction from sparse views.
• To enhance multi-view consistency in 3D Gaussian Splatting.
• To improve the robustness, efficiency, and memory consumption of 3D scene representation.

Main Methods:

• Leveraging sparse matching priors to initialize Gaussians, prioritizing textured regions and using random distribution for low-texture areas, creating a compact initial set.
• Implementing a multi-view consistency-guided progressive pruning strategy to dynamically remove inconsistent Gaussians.
• Confining Gaussian optimization within a consistency-constrained space for robust reconstruction.

Main Results:

• Achieved photorealistic scene reconstruction from sparse input views.
• Demonstrated enhanced robustness to sparse view conditions.
• Showcased accelerated rendering speeds and reduced memory consumption compared to existing methods.

Conclusions:

• MCGS offers a practical framework for 3D Gaussian Splatting, effectively addressing challenges posed by sparse input views.
• The proposed methods significantly improve multi-view consistency, leading to more reliable 3D scene representations.
• MCGS provides a balance of high-quality reconstruction, efficiency, and reduced resource usage.