DensifyBeforehand: LiDAR-assisted Content-aware Densification for Efficient and Quality 3D Gaussian Splatting

TL;DR

This work tackles inefficiencies and artifacts of adaptive density control in 3D Gaussian Splatting by introducing an offline densification pipeline that fuses sparse LiDAR data from a mobile device with monocular depth estimates to produce a dense, ROI-focused initialization. By applying ROI-aware sampling and avoiding the ADC stage during optimization, the method initializes a compact set of Gaussians that still achieves high visual fidelity. Experiments on six self-collected datasets show comparable or superior rendering quality to state-of-the-art baselines while significantly reducing the number of Gaussians and training time. The approach advances practical, resource-efficient 3D scene representations suitable for mobile sensing and large-scale reconstructions.

Abstract

This paper addresses the limitations of existing 3D Gaussian Splatting (3DGS) methods, particularly their reliance on adaptive density control, which can lead to floating artifacts and inefficient resource usage. We propose a novel densify beforehand approach that enhances the initialization of 3D scenes by combining sparse LiDAR data with monocular depth estimation from corresponding RGB images. Our ROI-aware sampling scheme prioritizes semantically and geometrically important regions, yielding a dense point cloud that improves visual fidelity and computational efficiency. This densify beforehand approach bypasses the adaptive density control that may introduce redundant Gaussians in the original pipeline, allowing the optimization to focus on the other attributes of 3D Gaussian primitives, reducing overlap while enhancing visual quality. Our method achieves comparable results to state-of-the-art techniques while significantly lowering resource consumption and training time. We validate our approach through extensive comparisons and ablation studies on four newly collected datasets, showcasing its effectiveness in preserving regions of interest in complex scenes.

Figures (5)

• Figure 1: Our performance comparing to other SOTA methods on wetlab dataset (a) and corner dataset (b). Our method can achieve comparable performance in terms of rendering quality and use fewer Gaussians to represent the same scene.
• Figure 2: Our "densify beforehand" approach. Our method takes as input a set of posed RGB frames real-time tracked from a visual-LiDAR device and the resulting sparse point cloud. We adopt a monocular depth estimation (MDE) method to derive dense depth maps from RGB frames and utilize the sparse LiDAR points to rescale the MDE results. To prepare dense initial points for training 3DGS, we adopt an ROI-aware importance sampling strategy. Finally we train the dense input points with 3DGS splitting and pruning to yield an efficient and compact scene.
• Figure 3: Qualitative comparisons of our results and those produced by the SOTA methods. Our method can capture the puller in the Wetlab scene (in the first row) and reconstruct both the text (enclosed) and the far-away scene of the Corner scene (in the second row). We achieve comparable visual quality with the comparing methods in the Pantry (third) and Staircase (fourth) where PixelGS excels in capturing the texts.
• Figure 4: Novel-view renderings to showcase floating artifacts. Left: LightGS, Right: Our result.
• Figure 5: User-specified semantic mask and improved readability of the poster in the Staircase scene.