Qmesh4point

Parameter-free and Training-free Deep Learning for Crown Reconstruction

1.Features

• High-accuracy crown volume estimation
• Parameter-free
• Automated 3D modeling from LiDAR/point clouds
• Cross-section analysis & projection mapping
• Low VRAM usage & fast processing
• Training-free for all tree species

2. Usage

Prerequisites

• Requires PyTorch ≥2.0 (versions below 2.0 not supported)
• Requires PyTorch3D ==0.7.5

Step 1: Convert XYZ to PLY

Run xyz2ply under the tools directory to convert your XYZ point cloud file (first three columns as XYZ coordinates) to a PLY file with normal vectors.

Step 2: Get Mesh Face Count

Execute get_mesh_number.py in tools/get_mesh_num/. You must modify the path to your target folder. Note this is a modified version of Adtree - if execution fails, please recompile. The original Adtree cannot retrieve face counts.

Step 3: Batch Convex Hull Processing

Run convex_hull_batch.py after:

1. Setting your target folder path
2. Specifying your face count CSV file path
3. Compiling Manifold (Robust Watertight Manifold Software) under the code directory

Step 4: Simulation Execution

Run simulate.py with these configurable options:

• ​​Multi-threading support​​: Set the number of parallel processes per device
• ​​Multi-GPU support​​: Configure the number of GPUs to utilize
• ​​Progress notifications​​: Enable and configure notification settings as needed

For optimal performance:

1. Adjust thread/GPU counts according to your hardware specifications
2. Allocate sufficient memory for large datasets
3. Monitor system resources during execution

3. Evaluation

Method	Real-world Dataset (3,000+ trees)	ForInstance-999 Dataset	Synthetic Dataset (Include True Label)
Point2Tree	✅smallest	✅ smallest	✅ Best
Voxel-based	-	-	-
Graham Slicing	-	-	-
Convex Hull	-	-	-

Key: ✅ = Our method performs better than the baseline approaches in both volume estimation and projection area calculation

Method	Volume Estimation	Projection Area
Point2Tree	✅ smallest	✅ smallest
Voxel-based	-	-
Graham Slicing	-	-
Convex Hull	-	-

Note: Demonstrated superior performance across all test datasets (3,000+ real trees, ForInstance-999, and synthetic Blender models) compared to traditional methods

4. Performance

I. GPU Memory Usage Comparison

Point2Tree reduces GPU memory consumption by 50% compared to Point2Mesh .

II. Processing Time Comparison

Point2Tree achieves 2× faster processing speed .

5. Citation

This work builds upon two foundational papers. We sincerely thank the authors for their outstanding contributions that made this research possible:

Qmesh4point Reference

@article{,
  title     = {QMesh4Point: Using Segformer and Self-Prior Learning to Calculate High-Precision Crown Volume and Projection Area Based on LiDAR Point Cloud Measurements},
  author    = {Zhong Wang,Zhixiang Xu,Ruishan Geng,Jiaxin Zhong Shan Yang},
  year      = {2025},
}

Point2Mesh Reference

@article{Hanocka2020p2m,
  title     = {Point2Mesh: A Self-Prior for Deformable Meshes},
  author    = {Hanocka, Rana and Metzer, Gal and Giryes, Raja and Cohen-Or, Daniel},
  journal   = {ACM Transactions on Graphics},
  volume    = {39},
  number    = {4},
  pages     = {Article 86},
  year      = {2020},
  doi       = {10.1145/3386569.3392415}
}

AdTree Reference

@article{du2019adtree,
  title     = {AdTree: Accurate, Detailed, and Automatic Modelling of Laser-Scanned Trees},
  author    = {Du, Shenglan and Lindenbergh, Roderik and Ledoux, Hugo and Stoter, Jantien and Nan, Liangliang},
  journal   = {Remote Sensing},
  volume    = {11},
  number    = {18},
  pages     = {2074},
  year      = {2019},
  doi       = {10.3390/rs11182074}
}