Abstract: The field of computer vision has seen significant advancements in recent years, resulting in a range of practical applications, such as virtual and augmented reality and self-driving cars. However, the reliance on costly and high-quality human annotations can hinder complex and novel vision tasks. In this talk, I will discuss ways to achieve detailed visual recognition and 3D understanding from in-the-wild images and videos without human annotations. I will then introduce a general framework that can train deep models from heterogeneous labels. I will also review my work on designing and theoretically understanding neural network architectures in the context of unsupervised image restoration tasks. These methods not only address core computer vision challenges but also open up new opportunities in ecology, material discovery, and 3D content creation.
Advisor: Subhransu Maji