gbemi123/differentiable-ray-tracing-for-optical-parts

Master thesis for Computer Graphics, Vision and Imaging at the University College London (UCL).

Differentiable Ray tracing for Designing Optical Parts

This repository contains the code used for the MSc project done at the Computational Light Laboratory at University College London. It was supervised by Kaan Aksit.

Abstract

The goal of this project was to investigate the design of an optical part with differentiable ray tracing and modern machine learning techniques. The optical component is specifically designed as an augmented reality headset made up of pinholes of varying sizes and that is placed 1cm from an image source. The differentiability of the optical component is made possible through the PyTorch library, and the ray tracing aspect is sped up through a graphics processing unit called CUDA. The renderer is created with the help of the ODAK computer graphics and visual perception library.

The algorithm would render a visualisation of looking through the optical component when the pinholes are placed randomly and then use stochastic gradient descent to optimise the component until a clearer image is generated. The result of testing the system on 8 images gave an average difference of 2.8×10^−3 between the test image and the image that is seen when looking through the optimised aperture array component.

System Design