Micro-splatting: Maximizing Isotropic Constraints for Refined Optimization in 3D Gaussian Splatting

2025-04-08

AI Summary

Traditional 3D Gaussian Splatting struggles with fine-grained detail reconstruction in large-scale scenes, as conventional covariance optimization often induces blur or sparsity imbalance. Method: We propose isotropic covariance regularization coupled with an image-gradient-driven adaptive Gaussian splitting strategy—dynamically densifying Gaussians in high-gradient regions—and jointly optimize a multi-objective loss comprising L1, L2, PSNR, SSIM, and LPIPS. Contribution/Results: To our knowledge, this is the first work to enforce isotropic constraints on covariances to ensure geometric consistency, and to design a gradient-adaptive splitting threshold that jointly optimizes detail fidelity and rendering efficiency. Experiments on multiple standard benchmarks demonstrate consistent improvements: PSNR and SSIM increase by over 1.5 dB, while LPIPS decreases by 12%, yielding significant enhancements in visual detail preservation and structural fidelity.

Technology Category

• Computer Vision: 3D Computer Vision
• Search and Optimization: Non-convex Optimization
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Security and Privacy: Large-scale security measurements
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Recent advancements in 3D Gaussian Splatting have achieved impressive scalability and real-time rendering for large-scale scenes but often fall short in capturing fine-grained details. Conventional approaches that rely on relatively large covariance parameters tend to produce blurred representations, while directly reducing covariance sizes leads to sparsity. In this work, we introduce Micro-splatting (Maximizing Isotropic Constraints for Refined Optimization in 3D Gaussian Splatting), a novel framework designed to overcome these limitations. Our approach leverages a covariance regularization term to penalize excessively large Gaussians to ensure each splat remains compact and isotropic. This work implements an adaptive densification strategy that dynamically refines regions with high image gradients by lowering the splitting threshold, followed by loss function enhancement. This strategy results in a denser and more detailed gaussian means where needed, without sacrificing rendering efficiency. Quantitative evaluations using metrics such as L1, L2, PSNR, SSIM, and LPIPS, alongside qualitative comparisons demonstrate that our method significantly enhances fine-details in 3D reconstructions.

Problem

• Improves fine-grained detail in 3D Gaussian Splatting
• Reduces blurring and sparsity in 3D representations
• Enhances rendering quality without sacrificing efficiency

Innovation

• Covariance regularization for compact isotropic splats
• Adaptive densification in high gradient regions
• Loss function enhancement for detailed reconstructions