WAMnet

Motivation

Allow me to introduce “WAMnet,” or “Wait-a-minute network.”

This was a captivating project that we tackled as a group. We willingly put some constraints on ourselves and ventured to construct a denoiser that could function in a fragment shader! Unlike the typical neural networks that rely on CUDA and fully exploit the GPU power, our WAMnet operates right within a fragment shader.

The Project

The implementation isn’t complex at all. Provided an NxN patch from a noisy input image, as well as the related normal map, our network makes a guess at the color of the center pixel of that patch. This methodology lets the network consider the data from neighboring pixels, the normal maps and the inferred surface properties. WAMnet, our creation, is a 5-layer Multi-layer perceptron, each having hidden layers containing 32 neurons. These neurons are sparked by a periodic activation function, drawing inspiration from the SiREN network.

Note: The above paragraph is derived from the implementation details from our report.

The interesting part about the project is the demo that follows!