Recent studies showed that the structure of convolutional neural networks (CNNs) can capture significant image statistics prior to any learning, allowing one to accurately generate and restore an image using a randomly initialized CNN by early-stopping during training. This method is much more time and memory efficient than typical methods that train CNNs to learn realistic image priors from large datasets of input-output image pairs. This project focuses on extending that idea, called Deep Image Prior, to 3-D space. The goal is to design a deep CNN that uses a decoder network as an implicit prior to generate a voxel grid for a scene using only a few input images and no prior training. If successful, we can use the model to simplify 3-D image tasks like novel view synthesis and upsampling.

I am excited to delve deeper into the field of computer graphics and explore the applications of machine learning to image generation tasks. Through SuperUROP, I hope to apply the knowledge I have gained from my machine learning and algorithms classes to a meaningful, long-term research project. Additionally, I hope to learn how to identify research problems and communicate results effectively.