<a href="https://laka-3dv.github.io/">Shaocheng Yan</a> <a href="https://shipc-ai.github.io/">Pengcheng Shi</a><sup>†</sup> <a href="https://ericzzj1989.github.io/">Zhenjun Zhao</a> Kaixin Wang Kuang Cao Ji Wu <a href="https://ljy-rs.github.io/web/">Jiayuan Li</a><sup>†</sup>

<p align="center"> <a href="https://arxiv.org/abs/2507.01439">  <img src="./assets/pipeline2.svg" alt="Teaser" width=70%> </a> <br> <em>A highly efficient and robust estimator for point cloud registration (PCR), supporting both CPU and GPU platforms.</em> </p> </div>

🌀 Overview

TurboReg is a highly efficient and robust estimator for point cloud registration (PCR), supporting both CPU and GPU platforms. It achieves 60+ FPS while maintaining state-of-the-art registration recall.

Highlights

State-of-the-art performance on 3DMatch and KITTI datasets
Efficient runtime (60+ FPS)
Learning-free, parameter-friendly design

Here's a general comparison of TurboReg with existing methods:

Comparison with Existing Methods

| Method Name | Venue (Year) | Performance (Recall) | CPU Efficiency | GPU Efficiency | |------------------|------------------|:------------------------:|:------------------:|:------------------:| | SC2-PCR | CVPR 2022 | ✅ | ❌ | ✅ | | MAC | CVPR 2023 | ✅ | ❌ | ❌ | | FastMAC | CVPR 2024 | ❌ | ✅ | ✅ | | TurboReg 🎖 | ICCV 2025 | ✅ | ✅ | ✅ (60+FPS) |

</div>

📦 Installation

Python API

Note: The Python API of TurboReg is implemented in C++ and exposed via PyBind11, offering C++-level performance within Python.

Create a Python virtual environment and install PyTorch (with CUDA support):

# Python Enviroment
git clone https://github.com/Laka-3DV/TurboReg.git
cd TurboReg

conda create -n turboreg python=3.11
conda activate turboreg

pip install open3d tyro

# Install PyTorch (select the appropriate version for your CUDA setup from https://pytorch.org/)
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Install TurboReg by running:

cd bindings
pip install .
cd ..

You can try TurboReg with just a few lines of code:

import numpy as np
import torch
import turboreg_gpu  # NOTE: torch must be imported before turboreg_gpu

# Load input correspondences
kpts_src = torch.from_numpy(np.loadtxt('demo_data/000_fpfh_kpts_src.txt')).cuda().float()
kpts_dst = torch.from_numpy(np.loadtxt('demo_data/000_fpfh_kpts_dst.txt')).cuda().float()

# Initialize TurboReg with specific parameters:
reger = turboreg_gpu.TurboRegGPU(
    6000,      # max_N: Maximum number of correspondences
    0.012,     # tau_length_consis: \tau (consistency threshold for feature length/distance)
    2000,      # num_pivot: Number of pivot points, K_1
    0.15,      # radiu_nms: Radius for avoiding the instability of the solution
    0.1,       # tau_inlier: Threshold for inlier points. NOTE: just for post-refinement (REF@PointDSC/SC2PCR/MAC)
    "IN"       # eval_metric: MetricType (e.g., "IN" for Inlier Number, or "MAE" / "MSE")
)

# Run registration
trans = reger.run_reg(kpts_src, kpts_dst)

To run the demo:

cd ./demo_py
python o3d_fpfh.py

C++ Build Instructions

🔧 Dependencies

CMake ≥ 3.10
LibTorch (No independent installation required; use directly within your Python virtual environment (i.e., turboreg))

GPU version uses LibTorch from the Python virtual environment. For CPU version, you can refer to ./turboreg_cpu/ (TBD).

mkdir build_ && cd build_ 
cmake .. -DCMAKE_PREFIX_PATH=`python -c "import torch; print(torch.utils.cmake_prefix_path)"`
make -j8

# Run C++ demo
./demo_cpp/demo_corr_fpfh

🧪 Experiment Reproduction

Download the processed 3DMatch correspondences from MAC++. Expected folder structure:

DIR_3DMATCH_FPFH_FCGF
.
├── 3DMatch
│   ├── all_fcgf
│       ├── 7-scenes-redkitchen
│       └── ...
│   └── all_fpfh
│       ├── 7-scenes-redkitchen
│       └── ...

Then, run the folloing script

# For 3DMatch + FPFH
python -m demo_py.exp_3dmatch --desc fpfh --dataname 3DMatch --dir_dataset "DIR_3DMATCH_FPFH_FCGF" --max_N 7000 --tau_length_consis 0.012 --num_pivot 2000 --radiu_nms 0.15 --tau_inlier 0.1 --metric_str "IN"

# For 3DMatch + FCGF
python -m demo_py.exp_3dmatch --desc fcgf --dataname 3DMatch --dir_dataset "DIR_3DMATCH_FPFH_FCGF" --max_N 6000 --tau_length_consis 0.012 --num_pivot 2000 --radiu_nms 0.10 --tau_inlier 0.1 --metric_str "MAE"

📈 Registration Performance (RR ↑, RE ↓, TE ↓, FPS ↑)

| Dataset | Descriptor | RR (%) | RE (°) | TE (cm) | FPS | | --------- | ---------- | ------ | ------ | ------- | ----- | | 3DMatch | FPFH | 84.10 | 2.19 | 6.81 | 52.87 | | 3DMatch | FCGF | 93.59 | 2.04 | 6.42 | 64.55 | | 3DMatch | Predator | 94.89 | 2.07 | 6.04 | 63.31 | | 3DLoMatch | FPFH | 40.99 | 3.85 | 10.16 | 62.31 | | 3DLoMatch | FCGF | 59.74 | 3.76 | 10.40 | 61.43 | | 3DLoMatch | Predator | 73.07 | 3.28 | 9.53 | 67.68 | | KITTI | FPFH | 98.56 | 0.47 | 8.96 | 68.93 | | KITTI | FCGF | 98.38 | 0.40 | 8.12 | 67.81 |

</div>

📖 Citation

If you find this code useful for your work or use it in your project, please consider citing:

@article{yan2025turboreg,
  title={TurboReg: TurboClique for Robust and Efficient Point Cloud Registration},
  author={Yan, Shaocheng and Shi, Pengcheng and Zhao, Zhenjun and Wang, Kaixin and Cao, Kuang and Wu, Ji and Li, Jiayuan},
  journal={arXiv preprint arXiv:2507.01439},
  year={2025}
}

🙏 Acknowledgements

TurboReg builds upon several excellent open-source works:

TurboReg

Install / Use

README