HyperGCT: A Dynamic Hyper-GNN-Learned Geometric Constraint for 3D Registration (ICCV 2025)

Source code of HyperGCT.

Introduction

Geometric constraints between feature matches are critical in 3D point cloud registration problems. Existing approaches typically model unordered matches as a consistency graph and sample consistent matches to generate hypotheses. However, explicit graph construction introduces noise, posing great challenges for handcrafted geometric constraints to render consistency. To overcome this, we propose HyperGCT, a flexible dynamic Hyper-GNN-learned geometric ConstrainT that leverages high-order consistency among 3D correspondences. To our knowledge, HyperGCT is the first method that mines robust geometric constraints from dynamic hypergraphs for 3D registration. By dynamically optimizing the hypergraph through vertex and edge feature aggregation, HyperGCT effectively captures the correlations among correspondences, leading to accurate hypothesis generation. Extensive experiments on 3DMatch, 3DLoMatch, KITTI-LC, and ETH show that HyperGCT achieves state-of-the-art performance. Furthermore, HyperGCT is robust to graph noise, demonstrating a significant advantage in terms of generalization.

Requirements

All experiments are conducted on RTX 3090 GPUs configured with CUDA 11.8. So CUDA and conda should be installed first, then you may configure HyperGCT as:

conda env create -f environment.yml
conda activate HyperGCT

Pretrained Model

We provide the pre-trained model of 3DMatch in snapshot/HyperGCT_3DMatch_release and KITTI in snapshot/HyperGCT_KITTI_release.

Data Preparation

The processed train and test sets can be found Here.

3DMatch / 3DLoMatch

We follow PointDSC to generate datasets. The data should be organized as follows:

--data--3DMatch                
        ├── fragments                 
        │   ├── 7-scene-redkitechen/
        |   |   ├── cloud_bin_0.ply
        |   |   ├── cloud_bin_0_fcgf.npz
        |   |   ├── cloud_bin_0_fpfh.npz
        │   |   └── ...      
        │   ├── 7-scenes-redkitchen_3dlomatch/
        │   |   └── ...        
        │   └── ...                
        └── threedmatch_feat              
            ├── 7-scenes-chess@seq-01_000_fcgf.npz
            ├── 7-scenes-chess@seq-01_001_fcgf.npz
            └── ...                               

KITTI-10m / KITTI-LC

We follow LiDAR Registration Benchmark to generate datasets. The data should be organized as follows:

--data--KITTI                
        ├── fcgf_train
        |   ├── drive0-pair0_13.npz
        |   ├── drive0-pair14_25.npz
        │   └── ...  
        ├── fcgf_val
        │   └── ...  
        ├── fcgf_test
        │   └── ...  
        ├── fpfh_test
        │   └── ...  
        ├── lc_fpfh_test_0_10
        |   ├── drive0-pair0_4438.npz
        |   ├── drive0-pair3_4432.npz
        │   └── ...  
        ├── lc_fpfh_test_10_20
        │   └── ...  
        └── lc_fpfh_test_20_30
            └── ...  

Instructions for training and testing

Before training and testing, modify the dataset path through config.root in the following scripts.

3DMatch

The training and testing on 3DMatch dataset can be done by running

# train on 3DMatch
python train_3dmatch.py
# test on 3DMatch
python test_3DMatch.py --chosen_snapshot [exp_id] --descriptor FCGF --use_icp False
# test on 3DLoMatch
python test_3DLoMatch.py --chosen_snapshot [exp_id] --descriptor FCGF --use_icp False

where the exp_id should be replaced by the snapshot folder name for testing (e.g., HyperGCT_3DMatch_release). The testing results will be saved in logs/.

KITTI

Similarly, the training and testing of KITTI dataset can be done by running

# train on KITTI-10m
python train_KITTI.py
# test on KITTI-10m
python test_KITTI.py --chosen_snapshot [exp_id] --descriptor FPFH --use_icp False
# test on KITTI-LC
python test_KITTI.py --chosen_snapshot [exp_id] --dataset KITTI_LC --range 0_10 --use_icp False

Acknowledgments

We thank the authors of

• PointDSC
• SC2PCR
• MAC
• VBReg

for open-sourcing their methods.

Citation

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

@InProceedings{Zhang_2025_ICCV,
    author    = {Zhang, Xiyu and Ma, Jiayi and Guo, Jianwei and Hu, Wei and Qi, Zhaoshuai and Hui, Fei and Yang, Jiaqi and Zhang, Yanning},
    title     = {HyperGCT: A Dynamic Hyper-GNN-Learned Geometric Constraint for 3D Registration},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {24750-24759}
}