Effective Gaussian Management for High-fidelity Scene Reconstruction.

Area of Science:

• Computer Vision
• 3D Graphics
• Geometric Modeling

Background:

• Current Gaussian Splatting (GS) methods optimize all primitives uniformly.
• This uniform treatment can lead to suboptimal performance in appearance and geometry reconstruction.
• Managing Gaussian attributes explicitly is crucial for enhanced scene representation.

Purpose of the Study:

• To propose an effective Gaussian management framework for high-fidelity scene reconstruction.
• To explicitly manage attribute activation, representation, and pruning of Gaussian primitives.
• To improve the efficiency and quality of 3D scene reconstruction using Gaussian-based methods.

Main Methods:

• Introduced GauSep: a densification strategy for selective attribute activation (color/normal) to mitigate gradient conflicts.
• Proposed GauRep: an adaptive Gaussian representation with dynamic spherical harmonics (SHs) orders and task-decoupled pruning.
• Developed CoRe: a regularized surface reconstruction module to distill robust normal fields from an SDF branch.

Main Results:

• The proposed framework achieves superior or comparable performance in appearance and geometry reconstruction.
• Significantly reduces the number of parameters compared to state-of-the-art methods.
• Demonstrates compatibility with various reconstruction architectures and seamless integration.

Conclusions:

• The Gaussian management framework effectively enhances high-fidelity scene reconstruction.
• Explicit attribute management leads to improved performance and reduced model size.
• The approach offers a versatile solution for advanced 3D scene representation.