MnRI Seminar: Hyun Soo Park
Neural Human Geometry: Reconstructing Humans from a Single Image
The geometry of humans is extremely complex, e.g., diverse articulated poses, stranded hairs, fine wrinkles, and stylistic dresses. Precise modeling of human geometry can bring realism to authentic social telepresence (through VR/AR devices). However, due to its complexity, it requires a large sensing infrastructure such as light field stages and multi-camera systems, which significantly limits real-world deployment.
Is it possible to model the geometry without such massive infrastructure, for example, with your cellphone? In this talk, I will address this question by presenting our ongoing effort on human geometry modeling from a monocular camera. A key challenge of making use of a monocular camera is that there is no geometric constraint, resulting in a fundamentally ill-posed problem. We tackle this challenge using deep neural networks that are designed to learn the plausible geometry and appearance of humans from a dataset called HUMBI--a large corpus of 3D models of human expressions from 772 individuals. With this dataset, I will showcase high fidelity 3D reconstruction of human geometry (depth, normal, texture).
Hyun Soo Park is an Assistant Professor at the Department of Computer Science and Engineering, the University of Minnesota (UMN). He is interested in computer vision approaches for behavioral imaging. He has received NSF's CRII and CAREER awards. Prior to UMN, he was a Postdoctoral Fellow in GRASP Lab at University of Pennsylvania. He earned his Ph.D. from Carnegie Mellon University.