INCODE: Implicit Neural Conditioning with Prior Knowledge Embeddings
_{^{WACV 2024}}

Project Page | Paper | Data

INCODE is a new method that improves Implicit Neural Representations (INRs) by dynamically adjusting activation functions using deep prior knowledge. Specifically, INCODE comprises a harmonizer network and a composer network, where the harmonizer network dynamically adjusts key parameters of the composer's activation function. It excels in signal representation, handles various tasks such as audio, image, and 3D reconstructions, and tackles complex challenges like neural radiance fields (NeRFs) and inverse problems (denoising, super-resolution, inpainting, CT reconstruction).

Amirhossein Kazerouni, Reza Azad, Alireza Hosseini, Dorit Merhof, Ulas Bagci

💥 News 💥

• 30.10.2023 | Code is released!
• 24.10.2023 | Accepted in WACV 2024! 🥳

Get started

Data

You can download the data utilized in the paper from this link.

Requirements

Install the requirements with:

pip install -r requirements.txt

Image Representation

The image experiment can be reproduced by running the train_image.ipynb notebook.

Audio Representation

The audio experiment can be reproduced by running the train_audio.ipynb notebook.

Shape Representation

The shape experiment can be reproduced by running the train_sdf.ipynb notebook. For your convenience, we have included the occupancy volume of Lucy with regular sampling in 512x512x512 cubes in the data file.

To test the model with custom input data, you can run the preprocess_sdf.ipynb notebook, which will generate a pre-processed .npy file for your desired input.
The output is a .dae file that can be visualized using software such as Meshlab (a cross-platform visualizer and editor for 3D models).

Image Denoising

The denoising experiment can be reproduced by running the train_denoising.ipynb notebook.

Image Super-resolution

The super-resolution experiment can be reproduced by running the train_sr.ipynb notebook.

CT Reconstruction

The CT reconstruction experiment can be reproduced by running the train_ct_reconstruction.ipynb notebook.

Image Inpainting

The inpainting experiment can be reproduced by running the train_inpainting.ipynb notebook.

Documentation

If you would like to replace the INCODE with other methods, including SIREN, FINER, Gauss, ReLU, SIREN, WIRE, WIRE2D,FFN, MFN, please refer to the Readme in the documentation folder.

Acknowledgement

We thank the authors of WIRE, MINER_pl, torch-ngp, and SIREN for inpainting for their code repositories.

Citation

@inproceedings{kazerouni2024incode,
  title={INCODE: Implicit Neural Conditioning with Prior Knowledge Embeddings},
  author={Kazerouni, Amirhossein and Azad, Reza and Hosseini, Alireza and Merhof, Dorit and Bagci, Ulas},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={1298--1307},
  year={2024}
}