STAF: Sinusoidal Trainable Activation Functions for Implicit Neural Representation

---

STAF is a novel approach that enhances Implicit Neural Representations (INRs) by introducing trainable sinusoidal activation functions. Specifically, STAF dynamically modulates its frequency components, enabling networks to adaptively learn and represent complex signals with higher precision and efficiency. It excels in signal representation, handling various tasks such as image, shape, and audio reconstructions, and tackles complex challenges like spectral bias and inverse problems, outperforming state-of-the-art methods in accuracy and reconstruction fidelity.

Get started

Data

You can download the data utilized in the paper from this link. Unzip the dataset, then copy it in the data directory in the code main directory.

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, Thai, Armadillo and Dragon in the data file.

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.

NTK

The NTK eigenfunctions and eigenvalues can be reproduced by running the figure_5.py of inr_dictionaries.