SCONE

This repository serves as the official code release of the AAAI24 paper: Learning Spatially Collaged Fourier Bases for Implicit Neural Representation

Jason Chun Lok Li^*, Chang Liu^*, Binxiao Huang, Ngai Wong

^*Contributed Equally

Department of Electrical and Electronic Engineering, The University of Hong Kong

⚙️ Dependency

conda create -n scone python=3.9
conda activate scone
pip install -r requirements.txt

🏗️ Code Structure

The repository contains training scripts train_<image/video/sdf>.py for various data modalities (image, video, SDF) as described in our paper. For convenience, we provide bash scripts in the scripts/ directory for quick start. Configuration files, including model and experiment settings, are stored as .yaml files under the config/ directory.

🧪 Experiments

Image

The Kodak dataset can be downloaded from this link. After downloading, please place the dataset in the data/kodak directory. To select which model to experiment, you can modify the model_config argument in the train_image.sh script. To train the model on all Kodak images in a single run, execute the following command in your terminal:

./scripts/train_image.sh

Video

The original cat video is available here. We have prepared for you the downsampled cat.npy file, which can be found in this link. Place it under the data/ folder. Once the data is ready, you can train the model on the cat video by executing the following command in your terminal:

./scripts/train_video.sh

SDF

The Stanford 3D scan dataset is available here. Download the .xyz files and place them in the data/stanford3d/ directory. Then, execute the command to start training on SDF data:

./scripts/train_sdf.sh

📝Citation

If you find SCONE is useful for your research and applications, consider citing it with the following BibTeX:

@inproceedings{li2024learning,
  title={Learning Spatially Collaged Fourier Bases for Implicit Neural Representation},
  author={Li, Jason Chun Lok and Liu, Chang and Huang, Binxiao and Wong, Ngai},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={12},
  pages={13492--13499},
  year={2024}
}

🙏🏼Acknowledgements

We have adapted some of our code from COIN++ and BACON. We sincerely thank them for their contributions to open source.