LiDAR to Mesh using Neural Kernel Surface Reconstruction

Related work

• Neural Kernel Surface Reconstruction
• fVDB
• PCD Meshing
• POCO

Setup

Repository

Clone the repository:

git clone git@github.com:dav1dclara/3d-vision.git
cd 3d-vision/

Dependencies

Create a conda environment, then install the dependencies:

conda create -n 3DV --file requirements.txt
conda activate 3DV
echo "/work/courses/3dv/team13/miniconda13/envs/nksr/lib/python3.10/site-packages/" > ~/miniconda3/envs/3dv/lib/python3.10/site-packages/shared_env.pth

Install the pre-commit hooks for automatic code formatting and linting on each commit:

pre-commit install

Reconstruction Pipeline

The reconstruction pipeline converts a LiDAR point cloud into a colored triangle mesh using a geometry-aware chunking strategy combined with Neural Kernel Surface Reconstruction (NKSR).

Overview

pointcloud.las + pointcloud.ply
        │
        ▼
Fine voxelization (0.5m cells)
        │
        ▼
Per-cell planarity analysis (SVD)
        │
        ▼
Region growing
  ├── Complex regions grow first, absorbing all adjacent planar chunks
  └── Remaining planar chunks form isolated planar-only regions
        │
        ▼
NKSR reconstruction per region
  ├── Complex units → high detail (vox=0.05m, detail=1.0, mise=2)
  └── Planar-only units → lower detail (vox=0.15-0.2m, detail=0.3-0.5)
        │
        ▼
Merge all chunks → reconstruction.ply

Key design decisions

Geometry-aware chunking — instead of splitting the scene into a regular grid (which cuts through surfaces and creates boundary artifacts), the scene is first split into uniform 0.5m cells, then each cell's planarity is estimated using PCA/SVD. The ratio of the smallest to largest singular value gives a residual — low residual means flat, high residual means complex geometry.

Complex-first region growing — complex regions (objects, furniture, people) grow first and greedily absorb all neighbouring planar chunks (walls, floors, ceilings). This means NKSR sees the full context of every edge — the wall and the object in the same reconstruction call — producing clean transitions without stitching artifacts. Only planar chunks that are never touched by a complex region form their own isolated planar reconstruction units.

Adaptive quality — complex units are reconstructed at the highest detail level with a fine voxel size (0.05m). Planar-only units use coarser voxels (0.15-0.2m) since flat surfaces reconstruct well at lower resolution, saving significant GPU time.

Normals from PLY — the NavVis scanner stores precomputed normals directly in the PLY file, which are used directly instead of estimating them from scratch (saving several minutes per run).

Configuration

All parameters are set in config.yaml

Running

conda activate 3DV
python scripts/nksr_reconstruction.py

Output is written to outputs/reconstruction.ply. Individual chunk PLYs are saved to outputs/chunks/ during reconstruction and can be inspected separately.

Output

The pipeline produces a colored triangle mesh in PLY format with per-vertex RGB colors projected from the original LiDAR scan. The mesh can be viewed in CloudCompare, MeshLab, or using the provided viewer script:

python scripts/view_mesh.py outputs/reconstruction.ply