TCF (Two-stage Consensus Filtering)

TCF is a 3D correspondence-based point cloud registration method. It elevates RANSAC to SOTA speed and accuracy. Firstly, one-point RANSAC obtains a consensus set based on length consistency. Subsequently, two-point RANSAC refines the set via angle consistency. Then, three-point RANSAC computes a coarse pose and removes outliers based on transformed correspondence’s distances. Drawing on optimizations from one-point and two-point RANSAC, three-point RANSAC requires only a few iterations. Eventually, an iterative reweighted least squares (IRLS) is applied to yield the optimal pose.

(a) Overall framework

(b) Outlier Removal

News

• 2025.01: The preprocessed dataset in TCF is available: Dataset Available.
• 2024.11: TCF has been accepted for IEEE Robotics and Automation Letters: Published Paper.

Citation

If you use code or data of TCF in your academic research, please cite our paper:

@ARTICLE{10758239,
  author={Shi, Pengcheng and Yan, Shaocheng and Xiao, Yilin and Liu, Xinyi and Zhang, Yongjun and Li, Jiayuan},
  journal={IEEE Robotics and Automation Letters}, 
  title={RANSAC Back to SOTA: A Two-Stage Consensus Filtering for Real-Time 3D Registration}, 
  year={2024},
  volume={9},
  number={12},
  pages={11881-11888},
  doi={10.1109/LRA.2024.3502056}}

Dependencies

• CMake
• PCL (Point Cloud Library)
• Eigen ≥ 3.4 (for slicing matrix)
• nlohmann_json (for reading config files)

File and data structure

Configure files

config
│   ├── config_eth.json
│   └── config_kitti.json

Data structure

data
├── ETH_TLS
│   └── facade
│       ├── s1_v0.1_sor.pcd
│       ├── s2_v0.1_sor.pcd
│       ├── s2_s1_top3.match
│       ├── s2-s1.pose
│       └── ...
├── KITTI
│   └── 09
│       ├── 210_v0.3.pcd
│       ├── 191_v0.3.pcd
│       ├── 210_191_top3.match
│       ├── 210-191.pose
│       └── ...

• xx.json: Contains paths for source point cloud, target point cloud, initial correspondences, and ground truth pose.
• xx.pcd: Source and target point cloud.
• A_B_top3.match: A correspondence file where A is source (left 3 columns) and B is target (right 3 columns).
• xx.pose: The 4*4 groundtruth pose from A to B

Note: Source and target point clouds are used to calculate resolution and estimate noise level, enabling a genralized method without manual parameter adjustments. If not available, values can be set manually, but correspondences remain necessary.

How to use?

Complie and run Demo

mkdir build && cd build
cmake ..
make -j8
./demo

Run other data

To test with different data, simply update the data path in the config file.