Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians

TL;DR

The paper tackles the inefficiency of representing complex scenes with millions of Gaussians in 3D Gaussian Splatting due to nonuniform spatial distribution. It introduces Mini-Splatting, a densification–simplification framework that repositions Gaussians using blur split and depth initialization, followed by intersection-preserving and sampling-based simplification, all within a rasterization pipeline. Three implementation variants are proposed to balance resource use, rendering quality, and storage: Mini-Splatting, Mini-Splatting-D, and Mini-Splatting-C. Experimental results across multiple real-world datasets show that Mini-Splatting achieves higher rendering quality with fewer Gaussians and demonstrates efficiency gains and storage advantages, establishing a strong baseline for Gaussian-Splatting-based work.

Abstract

In this study, we explore the challenge of efficiently representing scenes with a constrained number of Gaussians. Our analysis shifts from traditional graphics and 2D computer vision to the perspective of point clouds, highlighting the inefficient spatial distribution of Gaussian representation as a key limitation in model performance. To address this, we introduce strategies for densification including blur split and depth reinitialization, and simplification through intersection preserving and sampling. These techniques reorganize the spatial positions of the Gaussians, resulting in significant improvements across various datasets and benchmarks in terms of rendering quality, resource consumption, and storage compression. Our Mini-Splatting integrates seamlessly with the original rasterization pipeline, providing a strong baseline for future research in Gaussian-Splatting-based works. \href{https://github.com/fatPeter/mini-splatting}{Code is available}.

Figures (15)

• Figure 1: By reorganizing the spatial distribution of 3D Gaussians kerbl20233d, our Mini-Splatting reduces the number of Gaussians (in millions) while enhancing model performance in terms of rendering speed, training time, and rendering quality. The Gaussian centers of a foreground object bicycle are projected onto the rendered image as blue points, demonstrating the more uniform spatial distribution achieved by our algorithm.
• Figure 2: Visual analysis of Gaussian centers. (a): Visualization of the original 3DGS model's projected Gaussian centers, along with its corresponding rendering quality (PSNR in dB) and the number of Gaussians (in millions). Notably, we observe the phenomena of 'overlapping' and 'under-reconstruction'. (b), (c), (d), and (e): Visualization of Gaussian centers after applying different sampling techniques, including pruning, random sampling, grid sampling, and density-preserved sampling, respectively.
• Figure 3: Visual analysis of blur split. (a) Ground truth image. (b) Rendered image using the original 3DGS model kerbl20233d. (c) Rendered index of Gaussians with the maximum contribution to each pixel. (d) Rendered image after applying the blur split. (e) Rendered Gaussian index after applying the blur split.
• Figure 4: Visual analysis of depth map and depth points from 3DGS kerbl20233d. (a): Rendered depth map from alpha blending. (b): Reconstructed depth points from alpha blending with red boxes highlighting significant artifacts. (c): Reconstructed depth points from our depth map using Gaussian centers. (d): Reconstructed depth points from our depth map using mid-points.
• Figure 5: Depth points. (a): Depth points from our depth map using Gaussian centers. (b): Depth points from our depth map using mid-points. (c): Merged depth points from our depth map using mid-points.