PointCloudStitcher

Overview

PointCloudStitcher is a tool designed to stitch multiple .ply files into a single .ply file. It utilizes a high-performance point cloud registration algorithm, optimized for partial overlaps of 10–30%.

Requirements

Data

• .ply files must have at least 30% partial overlap. A higher overlap percentage generally results in a more accurately stitched final file.
• Files should be named in ascending or alphabetical order, ensuring that each file overlaps with the next one in sequence.
• Files should be located in ./data/experiments/exp/ply, relative to the repository directory.
• It's recommended to lightly clean the point clouds—e.g., by removing outliers or distant points in CloudCompare—to improve alignment and results.

Software

• CUDA: Version 11.X (Tested in 11.8 and 11.7)
• cuDNN: Version > 8.9 (Tested – Other versions may also work)

Installation

Follow these steps to install the software:

# Create a new virtual environment (recommended)
conda create -n pointcloud_stitcher python==3.10
conda activate pointcloud_stitcher

# Clone the repository
git clone https://github.com/djetshu/PointCloudStitcher.git # git@github.com:djetshu/PointCloudStitcher.git
cd PointCloudStitcher

# Install PyTorch 2.1 (Ensure compatibility with CUDA 11.8)
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118

# Install dependencies
pip install -r requirements.txt

# Build the package
python setup.py build develop

Pre-trained Weights

Download the geotransformer-3dmatch.pth.tar weight file from the release page.

After downloading, place the weights (geotransformer-3dmatch.pth.tar) in the weights/ directory.

Quick Test

Steps:

1. Ensure .ply files are located in ./data/experiments/exp/ply. If the exp or ply directories do not exist, create them.

Data Organization

The dataset should be structured as follows:

--data
   |--experiments
       |--exp
           |--ply
               |--rgb_0.ply
               |--rgb_1.ply
               |--rgb_2.ply
               |--rgb_3.ply
           |--npy (Created by the program)
           |--output (Created by the program)

2. Run the following command:

python pointcloud_stitching.py --data-path data/experiments/exp

Arguments:

• --data-path: Path to the experiment directory (default: data/experiments/exp).
• --weights: Path to the pre-trained model weights (default: ./weights/geotransformer-3dmatch.pth.tar).
• --no-graph: Disables graphical visualization of the stitching process.
• --output: Directory for saving the final .ply file (default: data/experiments/exp/output).

3. Once processing is complete, the final .ply files (stitched.ply and stitched_rgb.ply) will be located in data/experiments/exp/output.

Geometric Transformer for Fast and Robust Point Cloud Registration

PointCloudStitcher is based on GeoTransformer, an implementation of the paper:

Geometric Transformer for Fast and Robust Point Cloud Registration

Tested Environment:

• OS: Windows 11 (Branch: main) and Ubuntu 22.04 (Branch: dev/pointcloud_stitching)
• Python: 3.10
• PyTorch: 2.1.2
• CUDA: 11.8
• cuDNN: 8.9