Toward clean and efficient 3D Gaussian representations.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Deep Learning

Background:

• 3D Gaussian splatting (3DGS) excels at novel-view synthesis but suffers from geometric irregularities due to photometric-only optimization.
• This leads to outliers and overly dense Gaussian distributions, compromising model quality and efficiency.

Purpose of the Study:

• To introduce a novel, geometry-aware optimization pipeline for 3DGS that enhances geometric regularity and efficiency.
• To achieve a cleaner Gaussian representation without external supervision while maintaining reconstruction fidelity.

Main Methods:

• Initial Point Purification (IPP): Removes outliers from the seed point cloud using radial stratification and k-NN density estimation.
• Anchor-Constrained Densification (ACD): Regulates Gaussian spawning to prevent color-driven drift and uncontrolled growth.
• Opacity-Sparsity Joint Pruning (OSJP): Progressively removes redundant Gaussians based on joint opacity and density sparsity.

Main Results:

• Reduced training time by 32.6% and rank-weighted dispersion (RW-Disp) by 29.8%.
• Decreased 95th-percentile inter-point distance (P95) by 42.2% and symmetric Chamfer distance by 80.5%.
• Maintained comparable reconstruction quality, resulting in a more compact and cleaner Gaussian set.

Conclusions:

• The proposed geometry-aware pipeline effectively optimizes 3DGS for improved geometric cleanliness and efficiency.
• The method requires no pretrained models, integrates seamlessly with existing 3DGS frameworks, and enhances downstream usability.
• This offers a concise and effective approach for optimizing 3D representations in computer graphics and vision.