Building systems that can interpret and reason about the 3D world is a fundamental problem of computer vision research. In order to achieve this goal, these systems must first learn to encode the geometry and semantic content of the world through limited 2D observations. In this project we hope to explore methods that can learn scene representations of this kind in a completely self-supervised fashion. Specifically, our goal is to develop models that with few 2D image observations of a 3D scene can produce a representation which encodes useful information (semantic, geometric, etc.) while still modelling uncertainty accurately. Finally, we will design techniques for evaluating the effectiveness of these learned representations on a variety of downstream tasks.

Through this SuperUROP, I hope to continue my work with the CSAIL Scene Representation Group in developing vision systems that can understand and reason about the physical world. In the past few years, I have taken multiple machine learning classes and have completed two summer research internships at Google. With this preparation in mind, I am excited to take on this challenge in order to further my research and technical abilities.