Efficient Density Control for 3D Gaussian Splatting

TL;DR

The paper identifies inefficiencies in Adaptive Density Control (ADC) densification within 3D Gaussian Splatting (3DGS) and addresses them with two innovations: Long-Axis Split (LAS), which preserves geometry while densifying along the Gaussian's longest axis, and Recovery-Aware Pruning (RAP), which suppresses overfitted Gaussians by exploiting differential opacity recovery after resets. LAS reduces pre/post-split discrepancies and overlaps, while RAP improves generalization by pruning misfitting components without sacrificing training efficiency. Across real-world datasets (Mip-NeRF 360, Tanks & Temples, Deep Blending), the approach consistently yields higher PSNR with fewer Gaussians and faster rendering, outperforming baselines such as 3DGS and TamingGS variants. The work provides a practical, plug-and-play improvement to 3DGS that enhances rendering quality and efficiency for real-time novel view synthesis.

Abstract

3D Gaussian Splatting (3DGS) has demonstrated outstanding performance in novel view synthesis, achieving a balance between rendering quality and real-time performance. 3DGS employs Adaptive Density Control (ADC) to increase the number of Gaussians. However, the clone and split operations within ADC are not sufficiently efficient, impacting optimization speed and detail recovery. Additionally, overfitted Gaussians that affect rendering quality may exist, and the original ADC is unable to remove them. To address these issues, we propose two key innovations: (1) Long-Axis Split, which precisely controls the position, shape, and opacity of child Gaussians to minimize the difference before and after splitting. (2) Recovery-Aware Pruning, which leverages differences in recovery speed after resetting opacity to prune overfitted Gaussians, thereby improving generalization performance. Experimental results show that our method significantly enhances rendering quality. Due to resubmission reasons, this version has been abandoned. The improved version is available at https://xiaobin2001.github.io/improved-gs-web .

Figures (9)

• Figure 1: By improving adaptive density control of 3DGS kerbl20233d, our EDC achieves superior rendering quality while using less Gaussians. Benefiting from Long-Aixs Split, we successfully recover details of the ceiling lights and smoke detectors in the drjohnson hedman2018deep scene.
• Figure 2: Our proposed Long-Axis Split can minimize the differences before and after splitting, thereby improving the optimization speed post-split.
• Figure 3: Compared to the normal case (below), two highly overlapping Gaussians after cloning (above) are more difficult to individually optimize for precise scene fitting.
• Figure 4: Split uses probabilistic sampling to generate the coordinates of child Gaussians. When the initial positions of the two child Gaussians are more reasonable (below), the final fitting result is better. Conversely, when the initial positions are less reasonable (above), the final fitting result is poorer.
• Figure 5: The original split method does not alter the shape of the sub-bodies, resulting in a shape formed by the child Gaussians that differs from the original shape of the parent Gaussian. In contrast, our method shortens the child Gaussians along their longest axis, ensuring that the shape of the covered region before and after the split remains approximately the same, thereby maximizing the densification efficiency.