DNGaussian++: Improving Sparse-View Gaussian Radiance Fields With Depth Normalization.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Novel view synthesis from sparse inputs is challenging for radiance fields.
• Existing methods face high computational costs and limited refinement capabilities.
• 3D Gaussian Splatting shows promise but degrades geometry with fewer input views.

Purpose of the Study:

• To introduce DNGaussian, a depth-regularized framework for efficient and high-quality few-shot novel view synthesis.
• To address geometry degradation in 3D Gaussian Splatting under sparse view conditions.
• To propose DNGaussian++, an extended version for improved geometry consistency and representation efficiency.

Main Methods:

• Developed DNGaussian with Hard and Soft Depth Regularization for accurate scene geometry under monocular depth supervision.
• Introduced Global-Local Depth Normalization to refine detailed geometry.
• Proposed DNGaussian++ incorporating a Geometry Instance Regularizer and Depth-Guided Geometry Reorganization.

Main Results:

• DNGaussian offers real-time, high-quality few-shot novel view synthesis at low costs.
• DNGaussian++ demonstrates state-of-the-art performance across multiple datasets and scenarios.
• The extended framework shows improved geometry consistency and representation efficiency.

Conclusions:

• DNGaussian++ effectively overcomes limitations of previous methods in sparse-view novel view synthesis.
• The framework achieves superior performance and efficiency through advanced depth regularization techniques.
• DNGaussian++ proves broadly applicable and effective across various backbones and tasks.