This will be a hybrid event with in-person attendance in Raisler Lounge (Towne 225) and virtual attendance on Zoom.

ABSTRACT

Modern reconstruction systems, leveraging advances in pose estimation and generative modeling, can reconstruct everyday objects and scenes from casually captured images. In this talk, I will highlight recent work from our group that pushes these frontiers. I will first present a learning-based approach to the structure-from-motion task, highlighting how we can rethink camera parametrization for neural prediction methods. Building on this, I will describe a framework for combining generative priors and pose estimation for high-fidelity 3D reconstruction ‘in-the-wild’. Finally, I will outline an approach for ‘ultra-fast’ 3D generation that allows users to synthesize high-fidelity 3D assets in under a second.