Improving Adaptive Density Control for 3D Gaussian Splatting

TL;DR

This work improves adaptive density control for 3D Gaussian Splatting by introducing three core enhancements: Corrected Scene-Extent, an Exponentially Ascending Gradient Threshold, and Significance-aware Pruning. Built on PixelGS densification and opacity-correction, these measures reduce artifacts in under- and over-reconstructed regions while maintaining a comparable Gaussian count, and they yield substantially faster training. Quantitative and qualitative results across multiple datasets show higher reconstruction fidelity (PSNR/SSIM) and lower perceptual error (LPIPS), with the method achieving faster convergence—often outperforming 3DGS after only 15k iterations. The approach is compatible with existing 3DGS derivatives, offering a practical and efficient path to more accurate and compact 3D scene reconstructions for novel-view synthesis.

Abstract

3D Gaussian Splatting (3DGS) has become one of the most influential works in the past year. Due to its efficient and high-quality novel view synthesis capabilities, it has been widely adopted in many research fields and applications. Nevertheless, 3DGS still faces challenges to properly manage the number of Gaussian primitives that are used during scene reconstruction. Following the adaptive density control (ADC) mechanism of 3D Gaussian Splatting, new Gaussians in under-reconstructed regions are created, while Gaussians that do not contribute to the rendering quality are pruned. We observe that those criteria for densifying and pruning Gaussians can sometimes lead to worse rendering by introducing artifacts. We especially observe under-reconstructed background or overfitted foreground regions. To encounter both problems, we propose three new improvements to the adaptive density control mechanism. Those include a correction for the scene extent calculation that does not only rely on camera positions, an exponentially ascending gradient threshold to improve training convergence, and significance-aware pruning strategy to avoid background artifacts. With these adaptions, we show that the rendering quality improves while using the same number of Gaussians primitives. Furthermore, with our improvements, the training converges considerably faster, allowing for more than twice as fast training times while yielding better quality than 3DGS. Finally, our contributions are easily compatible with most existing derivative works of 3DGS making them relevant for future works.

Figures (9)

• Figure 1: Rendered novel views from the drjohnson, garden and flowers scene using 3DGS, PixelGS and our method. Our method produces better backgrounds, creates less artifacts and uses the same number of Gaussians as 3DGS.
• Figure 2: Camera locations of the Garden scene, the drjohnson scene and the truck scene. Here, the camera positions are closely distributed around the table in the center of the garden.
• Figure 3: Our proposed gradient threshold schedule over the densification interval compared against the number of Gaussians created by our model. The number of densified Gaussians clearly corresponds to our exponential schedule.
• Figure 4: Number of Gaussians during the training. With our proposed exponentially ascending gradient threshold, we produce many Gaussians at the beginning of the training and few at the end. Here, we trained with the Drjohnson scene.
• Figure 5: Selected novel view from the Garden scene trained with the baseline (3DGS) in two configurations, one with and the other without pruning.