Superpixel-guided Sampling for Compact 3D Gaussian Splatting

Myoung Gon Kim, SeungWon Jeong, Seohyeon Park, JungHyun Han

[Paper]

Superpixel-guided sampling	Sampled points	Optimized 3D Gaussians	Rendering result

BibTex

@inproceedings{kim2024superpixel,
  title={Superpixel-guided Sampling for Compact 3D Gaussian Splatting},
  author={Kim, Myoung Gon and Jeong, SeungWon and Park, Seohyeon and Han, JungHyun},
  booktitle={Proceedings of the 30th ACM Symposium on Virtual Reality Software and Technology},
  pages={1--15},
  year={2024}
}

---

0. Pre-requisite

• pytorch
• cv2
• numpy
• tqdm
• + all pre-requisite of 3DGS

1. Dataset Pre-processing

This project implementation is primarily designed with the Replica dataset.
Users must pre-process the dataset to ensure it matches the input format required for this project.

1.1 Dataset

• RGB dataset (.png by default)
• Depthmap dataset (.tiff by default)

1.2 Directories and files from COLMAP

• sparse/0
  • COLMAP creates this directory which contains 'cameras.bin', 'images.bin' and 'points3D.bin' (or as .txt)
  • Locate it (copy) at the dataset path.
• images.txt
  • This file is from COLMAP. Locate it (copy) at the dataset path.

1.3 File lists

• rgb_list.txt

  • This is list of rgb image paths. in this .txt file, all images needs to be sorted and listed.

    path/rgb_0.png
    path/rgb_1.png
    path/rgb_2.png
    ...
    path/rgb_899.png 
  

• depth_list.txt

  • This is list of depth image paths. in this .txt file, all images needs to be sorted and listed.

    path/depth_0000.tiff
    path/depth_0001.tiff
    path/depth_0002.tiff
    ...
    path/depth_0899.tiff
  

• Locate them at your dataset path

2. Run

python run.py --scene_path C:/your_path

3. Result

Check results at your_dataset_path/save/

4. Optimize (train)

• Rename save/input.ply as sparse/0/points3D.ply
• Train it using Original 3D Gaussian-Splatting.
  • skip the convert.py part
  • In arguments/__init__.py(of Original 3DGS), modify two parameters
    • self.densify_from_iter = 500
    • self.densify_until_iter = 1600
  • python train.py -s c:/your_path

Funding and Acknowledgments

This research was supported by the Ministry of Science and ICT,Korea, under the ICT Creative Consilience Program (IITP-2024-2020-0-01819), ITRC (Information Technology Research Center)Support Program (IITP-2024-2020-0-01460) and the grant no. 2020-0-00861.

Our research builds upon the foundation established by Gaussian-Splatting / (github) and Hertel Mehlhorn Convex Decomposition / (github)