GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024)
Paper: arXiv
Project Page: https://kevinhuangxf.github.io/GaussianMarker/
git clone https://kevinhuangxf.github.io/GaussianMarker/ --recursive
# create conda environment
conda create -n gaussianmarker python=3.8
conda activate gaussianmarker
# install pytorch
pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu121
# install submodules
pip install submodules/diff-gaussian-rasterization/
pip install submodules/simple-knn/
pip install -e submodules/FisherRF/diff/ -v
# install requirements
pip install -r requirements.txt
Please download the Blender, LLFF and MipNeRF360 datasets in the official websites.
You can train your 3DGS model using this repo, or use the official 3D Gaussian Splatting:
python train.py -s /path/to/dataset
We provide the pretrained HiDDeN model in the hidden folder based on the Stable Signature. You can run the pretrained HiDDeN model by:
PYTHONPATH=. python hidden/hidden_images.py hidden/imgs/ hidden/results
You can check the README.md file for how to train the HiDDeN model from scratch.
Train and evaluate the GaussianMarker on the 3DGS models:
# Train on Blender
python train_gaussianmarker.py -s /path/to/nerf_synthetic/lego -m path/to/3dgs_model --iterations 2000
# Train on LLFF and save visualization results
python train_gaussianmarker.py -s /path/to/nerf_llff_colmap/trex -m path/to/3dgs_model --iterations 1000 --test_iterations 1000 --save_iterations 1000 --save_vis
# Train on MipNeRF360 with 1/4 resolotion for saving memory and save visualization results
python train_gaussianmarker.py -s /path/to/mip360/bicycle -m path/to/3dgs_model --iterations 2000 --save_vis -r 4
Train 3D decoder (optional):
python train_gaussianmarker.py -m /path/to/nerf_synthetic/lego --train_3d_decoder --iterations 2000 --test_iterations 2000
@article{huang2024gaussianmarker,
title = {GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting},
author = {Xiufeng Huang, Ruiqi Li, Yiu-ming Cheung, Ka Chun Cheung, Simon See, Renjie Wan},
journal = {Neural Information Processing Systems (NeurIPS)},
year = {2024},
}
This project builds heavily on 3D Gaussian Splatting, FisherRF and Stable Signature. We thanks the authors for their excellent works! If you use our code, please also consider citing their papers as well.