GrassmannRegistration (ICCV25, highlight)

This repository contains the code for GrassmannRegistration, an official implementation for the paper Registration beyond Points: General Affine Subspace Alignment via Geodesic Distance on Grassmann Manifold, which is accepted to ICCV 2025. GrassmannRegistration leverages geodesic distance on Grassmann manifold for cost function, leading to better convergence and extension to global optimization algorithm.

[arXiv] [BibTex]

GrassmannRegistration Supports

• Line to line registration
• Line to plane registration
• Plane to plane registration

News

• 25.10.05 We released an inital version of GrassmannRegistration!

1. Dependencies

The code was tested on following dependencies:

• Ubuntu 20.04
• Ceres Solver (2.2.0)
• Eigen3
• pmc (Parallel Maximum Clique Library)

2. Install

git clone https://github.com/joomeok/GrassmannRegistration.git
cd GrassmannRegistration
git submodule update --init --recursive
mkdir build
cd build 
cmake .. && make -j4

3. Examples

For now, we provide readily executable example only for BnB solver. However, you can refine your initial pose obtained from another solver with refine_main.cpp. Refer to detailed implementation within it for your specific example.

3.1 Data Format

For current version, you need two input files (features with respect to model and data frames, respectively). The format is:

Line feature

$(Feature Num)
sp_x sp_y sp_z ep_x ep_y ep_z id
...

Plane feature

$(Feature Num)
p1_x p1_y p1_z p2_x p2_y p2_z p3_x p3_y p3_z id
...

For line feature, sp and ep denote start point and end point respectively, while p1, p2, and p3 denote 3 linearly independent points on plane feature. id contains the correspondence information.

3.2 RGB-D Odometry

Run:

./RGBD_odom.sh

Refer to toy_example.ipynb for stacking delta pose to plot trajectory. You can compare resulted trajectory with ground truth trajectory in /data folder.

3.3 Line to Plane Registration

Refer to data generation code within toy_example.ipynb. You can run registration with:

./GoL2P $(YourDirectory)/lines.txt $(YourDirectory)/plane.txt ../config.txt $(YourDirectory)/result.txt

3.4 Custom Data

You can solve your own registration problem by running one of three binary files: GoL2L, GoL2P, and GoP2P. If you want to solve your plane-to-plane registration, try this command:

./GoP2P $(YourDirectory)/model_plane.txt $(YourDirectory)/data_plane.txt ../config.txt $(YourDirectory)/result.txt

Acknowledgements

This project is based on code from yangjialong/Go-ICP.

BibTex

@INPROCEEDINGS{jhshin-2025-iccv,  
    AUTHOR = { Jaeho Shin and Hyeonjae Gil and Junwoo Jang and Maani Ghaffari and Ayoung Kim },  
    TITLE = { Registration beyond Points: General Affine Subspace Alignment via Geodesic Distance on Grassmann Manifold },  
    BOOKTITLE = { Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) },  
    YEAR = { 2025 },  
    MONTH = { October },  
    ADDRESS = { Honolulu, Hawaii }
}