Multi-View Point Cloud Registration and Reconstruction for High-Fidelity Object Modeling

The purpose of this repository is to establish a high-fidelity 3D reconstruction system based on multi-frame point cloud registration. The point cloud data is derived from various viewpoints around the target object, acquired through scanning. The data sources can be depth cameras or LiDAR (Light Detection and Ranging).

Paper

The article is currently undergoing peer review:

Reconstruction results

• Results on simulated data (rendering by Open3D):

  dragon	deer	vase

• Results on real-world data:

Statue	Sofa

Requirements

• Eigen3(3.3.4)
• OpenCV (>4.0)
• Open3D
• Teaser-pp

Preparation

• Install a feature point-based registration algorithm according to Teaser-pp 's guidance, to serve as the initial registration pose for our algorithm.

• Download data from GoogleDrive and place it in the data folder.

• build project:

  mkdir build && cd build

  cmake .. && make

Usage

1. Firstly, use Teaser-pp to generate the initial pose, which is based on a feature point matching method, allowing for a rough alignment of the point cloud sequence.

   ./teaser_coarse_align ../cfg/simrecon_params.yaml

2. Start pairwise and global registration.

   ./multi_way_align_sim ../cfg/simrecon_params.yaml

Simulation and Real-world Datasets

All dataset can be downloded at GoogleDrive.

The real-world data is automatically acquired through the omnidirectional collection platform we designed:

Simulation data is collected using a Kinect camera in the Gazebo platform:

Note.

The code for the Gazebo simulation data platform, along with more details of our method, will be published subsequently.