GaussReg: Fast 3D Registration with Gaussian Splatting (ECCV2024)

Project Page | Paper

---

Install

This implementation uses Python 3.8.0, torch1.7.1+cu110. We recommend to use conda to deploy the environment.

• Install with conda:

conda env create -f environment.yml
conda activate GaussReg
python setup.py build develop

---

Demo

To test on your own data, please arange the data with the following structure:

|-- scene_name
    |-- A
        |-- output
            |-- point_cloud
                |-- iteration_30000
                    |-- point_cloud.ply
            |-- cameras.json    
            |-- cfg_args
            |-- input.ply
    |-- B
        |-- output
            |-- point_cloud
                |-- iteration_30000
                    |-- point_cloud.ply
            |-- cameras.json    
            |-- cfg_args
            |-- input.ply

1. Coarse Registration

The pretrained model of coarse registration can be download here. Put the pretrained models under

weights/coarse_registration.pth.tar

To get the coarse registration between your scenes, please run the following command:

python experiments/geotransformer.gaussian_splatting.indoor/demo.py 
        --src_file scene_name/B/output/point_cloud/iteration_30000/point_cloud.ply
        --ref_file scene_name/A/output/point_cloud/iteration_30000/point_cloud.ply
        --weights weights/coarse_registration.pth.tar

The output visualization of gs point clouds and estimated_transform.npz will be saved under 'demo_outputs'.

2. Fine Registration

TODO

3. Gaussian Splatting Fusion

To merge these two GS models (transform B into the coordinate system of A), please run the following command:

python gs_fusion.py --root_path scene_name --transform_path demo_outputs/estimated_transform.npz

The output file will be saved under 'scene_name/fuse'.

---

Training and Test on the ScanNet-GSReg Dataset

0. Data Preparation

In our paper, we construct a dataset called ScanNet-GSReg, comprising 1379 scenes from the ScanNet dataset. And our GaussReg is trained on the ScanNetGSReg training set. The proposed ScanNet-GSReg dataset can be download here. The dataset is organized as follows:

|-- ScanNet-GSReg
    |-- train
        |-- scene0000_00
            |-- A
                |-- output
                    |-- point_cloud
                        |-- iteration_10000
                            |-- point_cloud.ply
                    |-- cameras.json    
                    |-- cfg_args
                    |-- input.ply
                |-- images
                |-- sparse
                |-- features.h5
                |-- global-feats-netvlad.h5
                |-- matches.h5
                |-- pairs-netvlad.txt
            |-- B   
            |-- log.txt
        |-- scene0000_01
        ...
    |-- test
        |-- scene0707_00
        |-- scene0708_00
        ...
    |-- train.pkl 
    |-- test.pkl 
    |-- test_transformations.npz

1. Coarse Registration

The code for Coarse Registration is in 'experiments/geotransformer.gaussian_splatting.indoor'. First, you should set "_C.data.dataset_root" as the dataset root path in config.py. Run the following code to train the Coarse Registration network:

python experiments/geotransformer.gaussian_splatting.indoor/trainval.py

Run the following code to test the Coarse Registration network on the ScanNet-GSReg dataset:

python experiments/geotransformer.gaussian_splatting.indoor/test.py

2. Fine Registration

TODO

---

Plan of Release

• Release ScanNet-GSReg dataset
• Release training and test code of coarse registration on the ScanNet-GSReg dataset
• Release the code of Guassian Splatting fusion and filtering
• Release training and test code of fine registration on the ScanNet-GSReg dataset

Citation

If you find our work is helpful, please cite

@article{chang2024gaussreg,
  title={Gaussreg: Fast 3d registration with gaussian splatting},
  author={Chang, Jiahao and Xu, Yinglin and Li, Yihao and Chen, Yuantao and Han, Xiaoguang},
  journal={arXiv preprint arXiv:2407.05254},
  year={2024}
}

Acknowledgements

This repository is based on some excellent works, such as geotransformer and Gaussian Splatting. Many thanks.