Neural SDF

Generate neural SDF from mesh

This is an implementation of a Implicit Neural Representations with Periodic Activation Functions in C++ and Vulkan. The SIREN (Sinusoidal Representation Networks) model enables the representation of complex signals and shapes using periodic activation functions.

Prerequisites

• C++ Compiler (C++17 or higher)
• Vulkan SDK

Building

make

Usage

To train a neural SDF, run the following command:

./neural_sdf --train -i <input.obj> -n <layers> -s <layer_size> -o <out_weights>

Note: The input mesh must be watertight and triangulated before training.

The weights are saved in a float array in sequence by layers ($A_1$, $b_1$, $A_2$, $b_2$, ..., $A_k$, $b_k$), with matrices stored row-wise. The binary file does not contain the size of the array, since it is uniquely determined by the network architecture. The input is a float[3] 3D point and the output is a scalar float distance to surface. Activation function for all layers is sin($w_0 \cdot W x$ + $b$), except the last layer, which has no activation. $w_0$ is a constant equal to 30.

To render using the trained weights, run:

./neural_sdf --render -m <path_to_weights> -n <layers> -s <layer_size>

Example usage:

./neural-sdf --render -m data/sdf1_weights.bin -n 2 -s 64

Rendered image will be saved in the output directory.

License

This project is licensed under the MIT License - see the LICENSE file for details.

References

SIREN: Implicit Neural Representations with Periodic Activation Functions - Link