DensifyBeforehand

LiDAR-assisted 3D Gaussian Splatting

DensifyBeforehand

Content-aware densification before optimization for compact, high-quality 3D scene reconstruction.

Phurtivilai Patt · Leyang Huang · Yinqiang Zhang · Yang Lei

Accepted to CVM 2026 as a Technical Brief

Paper PDF Code Demo data

Ground truth and novel-view renderings on Wetlab: original 3DGS, Pixel-GS, Taming 3DGS, and DensifyBeforehand.

Overview

Densify first.
Optimize what matters.

Adaptive density control grows Gaussian primitives during training, but can introduce redundant Gaussians and floating artifacts. DensifyBeforehand moves that work to initialization.

We combine sparse LiDAR from a mobile device with monocular depth estimation, refine depth globally and locally using projected LiDAR anchors, and sample a content-aware dense point cloud. The resulting initialization lets 3DGS optimize appearance and geometry without cloning, while retaining splitting and pruning.

The method preserves texture-rich regions and thin structures while using fewer final Gaussians and less training time across six newly collected scenes.

29.72
PSNR on Wetlab
252.9k
Final Gaussians
402.87s
Training time
5.52%
Low-opacity Gaussians

Method

Dense initialization from sparse mobile LiDAR

Posed RGB frames and a sparse LiDAR point cloud become a content-aware initialization before 3DGS optimization begins.

DensifyBeforehand pipeline: mobile RGB and LiDAR inputs, monocular depth estimation, LiDAR-assisted refinement, ROI-aware sampling, and 3DGS optimization.
The complete pipeline. Click the figure to inspect the full-resolution version.
  1. 01

    Refine depth

    Project visible LiDAR anchors into each view, then apply iterative global median scaling and nearest-anchor local correction to Metric3Dv2 depth.

  2. 02

    Sample by content

    Back-project RGB-variance importance into 3D and allocate samples at a 30:1 ROI-to-background ratio.

  3. 03

    Optimize compactly

    Train 3DGS without cloning. Splitting remains available, followed by a direct opacity prune at iteration 25,000.

Qualitative results

Details survive where sparse initialization struggles

Across Wetlab, Corner, Pantry, and Staircase, DensifyBeforehand preserves thin geometry and readable, texture-rich regions with a compact scene representation.

Qualitative comparison of ground truth, 3DGS, Pixel-GS, Taming 3DGS, LightGaussian, and DensifyBeforehand on four scenes.
Novel-view comparison. Green insets highlight thin structures, text, and foreground regions.

Wetlab benchmark

Quality with fewer Gaussians

All methods use the same LiDAR point cloud for initialization. Higher PSNR and SSIM are better; lower LPIPS, time, and Gaussian counts are better.

Method PSNR ↑ SSIM ↑ LPIPS ↓ Time (s) ↓ Final #G (k) ↓ Peak #G (k) ↓
3DGS28.85050.93870.1459407.61918.23918.23
Pixel-GS29.63150.93690.15002019.861439.351439.35
Taming 3DGS28.78540.93870.1456496.04406.86406.86
LightGaussian28.97270.93620.15802736.09285.42739.44
DensifyBeforehand29.71870.93890.1454402.87252.90630.71

Efficiency

A controlled Gaussian budget

The radar comparison summarizes quality, training time, final and peak point counts, and low-opacity Gaussians on Wetlab and Corner.

Radar charts comparing DensifyBeforehand with 3DGS, Pixel-GS, Taming 3DGS, and LightGaussian on quality and efficiency metrics.
Novel-view comparison showing floating artifacts above a workbench for LightGaussian and a cleaner DensifyBeforehand result.
Fewer floaters. Direct dense initialization avoids the yellow-highlighted floating artifacts produced by the comparison method.
Semantic region-of-interest mask and a reconstructed poster with improved readability.
User-directed detail. A semantic mask can redirect samples toward a selected object and improve text readability.

Runnable demo

Reproduce the Wetlab pipeline

The public Wetlab package contains 191 RGB frames, calibrated poses, 148,414 LiDAR points, and a reference initialization. Run preprocessing, clone-disabled training, rendering, and evaluation with the released code.

Download Wetlab View instructions
191
RGB frames
167 / 24
Train / test views
148,414
LiDAR points
300k
Point budget

Citation

Build on this work

Accepted to CVM 2026 as a Technical Brief. Please cite the arXiv paper when using the method, code, or Wetlab demonstration package.

@article{phurtivilai2025densifybeforehand,
  title   = {DensifyBeforehand: LiDAR-assisted Content-aware Densification
             for Efficient and Quality 3D Gaussian Splatting},
  author  = {Phurtivilai, Patt and Leyang, Huang and
             Yinqiang, Zhang and Yang, Lei},
  journal = {arXiv preprint arXiv:2511.19294},
  year    = {2025},
  doi     = {10.48550/arXiv.2511.19294}
}