StegaNeRF: Embedding Invisible Information within Neueral Radiance Fields. ICCV2023

Method

Quick start

Environment

. ./create_env.sh

Dataset

Please download the datasets from these links:

NeRF synthetic: Download nerf_synthetic.zip from https://drive.google.com/drive/folders/128yBriW1IG_3NJ5Rp7APSTZsJqdJdfc1
LLFF: Download nerf_llff_data.zip from https://drive.google.com/drive/folders/128yBriW1IG_3NJ5Rp7APSTZsJqdJdfc1

Training

cd opt && . ./stega_{llff/syn}.sh [scene_name] [embed_img]

At the first stage, a photorealistic radiance field will first be reconstructed if it doesn't exist on disk. Then the steganographic training at the second stage ends up with the steganographic NeRF and decoder.
Select {llff/syn} according to your data type. For example, use llff for flower scene, syn for lego scene.
[embed_img] is the style image inside ./data/watermarks.

Evaluation & Rendering

View the results by tensorboard.

You can also obtain the results and rendering the videos from the saved checkpoints.

Use opt/render_imgs.py for the scenes on LLFF: python render_imgs.py <CHECKPOINT.npz> <Decoder.pt> <data_dir>

Use opt/render_imgs_circle.py to render a spiral for the scenes on NeRF synthetic: python render_imgs_circle.py <CHECKPOINT.npz> <Decoder.pt> <data_dir>

Experiments on NeRF-W

Dataset: Download brandenburg_gate (4.0G) from https://www.cs.ubc.ca/~kmyi/imw2020/data.html. More details to use this dataset can be found here.

Acknowledgement

We would like to thank ARF and Plenoxel authors for open-sourcing their implementations.

Citation

If you find this repo is helpful, please consider citing:

@inproceedings{li2022steganerf,
        title={StegaNeRF: Embedding Invisible Information within Neural Radiance Fields},
        author={Chenxin Li and Brandon Y. Feng and Zhiwen Fan and Panwang Pan and Zhangyang Wang},
        booktitle={arxiv},
        year={2022}
      }

chenxinli001/StegaNeRF