Efficient Scene Modeling via Structure-Aware and Region-Prioritized 3D Gaussians

TL;DR

This work addresses inefficiencies in 3D Gaussian Splatting by introducing Mini-Splatting2, a geometry-guided framework that jointly enforces structure-aware Gaussian distributions and region-prioritized optimization. The method leverages adaptive Gaussian organization and redundant Gaussian simplification to achieve compact representations, and aggressive model growth plus occluded Gaussian culling to accelerate training while preserving quality. Across real-world benchmarks, Mini-Splatting2 attains up to $4\times$ fewer Gaussians and about $3\times$ faster optimization with state-of-the-art rendering fidelity, demonstrating practical gains for efficient 3D scene modeling. The approach establishes a foundation for geometry-aware, efficient 3D Gaussian modeling with broad implications for robotics, digital twins, and immersive content creation.

Abstract

Reconstructing 3D scenes with high fidelity and efficiency remains a central pursuit in computer vision and graphics. Recent advances in 3D Gaussian Splatting (3DGS) enable photorealistic rendering with Gaussian primitives, yet the modeling process remains governed predominantly by photometric supervision. This reliance often leads to irregular spatial distribution and indiscriminate primitive adjustments that largely ignore underlying geometric context. In this work, we rethink Gaussian modeling from a geometric standpoint and introduce Mini-Splatting2, an efficient scene modeling framework that couples structure-aware distribution and region-prioritized optimization, driving 3DGS into a geometry-regulated paradigm. The structure-aware distribution enforces spatial regularity through structured reorganization and representation sparsity, ensuring balanced structural coverage for compact organization. The region-prioritized optimization improves training discrimination through geometric saliency and computational selectivity, fostering appropriate structural emergence for fast convergence. These mechanisms alleviate the long-standing tension among representation compactness, convergence acceleration, and rendering fidelity. Extensive experiments demonstrate that Mini-Splatting2 achieves up to 4$\times$ fewer Gaussians and 3$\times$ faster optimization while maintaining state-of-the-art visual quality, paving the way towards structured and efficient 3D Gaussian modeling.

Figures (16)

• Figure 1: Gaussians in 3DGS exhibit a highly irregular spatial distribution, which inherently limits its representational capability. To address this, Mini-Splatting introduces a structure-aware distribution scheme that mitigates such irregularity, substantially reducing the number of Gaussians while improving rendering speed, training efficiency, and visual quality. Building on this foundation, Mini-Splatting2 incorporates a region-prioritized optimization scheme to further accelerate training. In a comparative evaluation on the bicycle scene, Mini-Splatting2 demonstrates substantial improvements over the latest official accelerated 3DGS implementation, referred to as 3DGS-accel kerbl20233dtaming3dgs, requiring 5.7$\times$ fewer Gaussians and achieving a 4.2$\times$ speedup in optimization, while maintaining state-of-the-art visual quality.
• Figure 2: Analysis of Gaussian Representation. (a) Projected Gaussian centers in the vanilla 3DGS, along with the corresponding rendering quality (PSNR in dB) and total number of Gaussians (in millions). Notable issues of 'overlapping' and 'under-reconstruction' are clearly observed. (b)-(e) Visualization of Gaussian centers after applying different simplification strategies, including pruning, random sampling, grid sampling, and density-preserved sampling, respectively.
• Figure 3: Spatial irregularity. (a, c) Ground-truth images of the playroom and bicycle scans. (b, d) Corresponding projected Gaussian centers, colored by density from low (blue) to high (red).
• Figure 4: Analysis of Gaussian Optimization. (a) Gaussian centers. (b) Corresponding depth points across 0.5K to 15K iterations. While the Gaussian centers exhibit some visual artifacts, a dense and informative point cloud can still be effectively extracted from the Gaussian representation.
• Figure 5: Visual analysis of the blur split component. (a) Ground truth image. (b) Image rendered using the original 3DGS kerbl20233d. (c) Gaussian indices corresponding to the maximum contribution for each pixel. (d) Image rendered after applying blur split. (e) Gaussian indices after applying blur split.