Micro-Splatting: Multistage Isotropy-Informed Covariance Regularization Optimization for High-Fidelity 3D Gaussian...

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• High-fidelity 3D Gaussian Splatting (3DGS) methods capture intricate details but suffer from large model sizes and lengthy training.
• Existing 3DGS approaches often require complex post-processing and auxiliary neural networks, increasing computational overhead.

Purpose of the Study:

• To develop a unified pipeline for high-fidelity 3D Gaussian Splatting that achieves significant model compactness without post-processing.
• To introduce novel regularization and adaptive densification techniques to optimize the Gaussian splatting process.

Main Methods:

• Introduced a trace-based covariance regularization in Stage I to promote near-isotropic Gaussians and improve color fitting.
• Implemented gradient-guided adaptive densification to selectively subdivide splats in complex regions, optimizing density.
• Developed a Stage II refinement process involving pruning low-impact splats and merging redundant neighbors using lightweight criteria.

Main Results:

• Achieved up to 60% reduction in splat count and model size across three benchmarks.
• Reduced training time by 20% compared to state-of-the-art methods.
• Maintained or surpassed state-of-the-art performance in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) metrics.

Conclusions:

• Micro-Splatting provides an efficient, end-to-end framework that balances high fidelity with model compactness in 3D Gaussian Splatting.
• The proposed method eliminates the need for post-processing and auxiliary modules, simplifying the 3DGS pipeline.
• Demonstrated the effectiveness of multistage optimization with isotropy-informed covariance regularization for compact and detailed 3D scene representations.