Supervisor

Suitable for

Abstract

Neural radiance fields have emerged as a promising means of reconstructing 3D scenes from multiple images, allowing photorealistic novel views of the scene to be rendered. However, because of the under-constrained nature of the reconstruction problem they suffer from artifacts such as “floaters” - where the neural field incorrectly places high density values in free space. The goal of this project is to develop a method to regularise the neural field and eliminate artifacts such as floaters. The key idea will be to use the inductive biases inherent to convolutional networks (see Deep Image Prior) to bias the reconstruction to a noise-free field. This should allow for improved quality with no additional training data.

Prerequisites: Suitable for those who have done the Machine Learning course. Having also done Computer Graphics would be useful.