ABSTRACT
Nature presents a captivating confluence of diversity and similarity. In order to make sense of our visual experiences in the world, humans as well as other natural intelligences are innately adept at recognizing the underlying intrinsic patterns, by simply looking at 2D projections of a constantly evolving 3D environment. Designing unsupervised perception systems to do the same is not only key to many AR/VR and robotics applications, but also a cornerstone for understanding visual perception in general. In pursuit of this ultimate goal, this talk will mainly focus on a recent line of effort in learning dynamic 3D objects like animals, simply from casuallyrecorded, unlabeled Internet images and videos. In addition, I will also briefly discuss a few other relevant works on inverse rendering, visual concept learning and spatial audio synthesis, which, collectively, attempt to explore the various aspects of our experiences in the natural world
Shangzhe Wu is a postdoc researcher at Stanford University, working with Jiajun Wu. He received his PhD from University of Oxford, advised by Andrea Vedaldi and Christian Rupprecht. His research focuses on unsupervised 3D perception and inverse rendering. He is particularly interested in uncovering the underlying statistical and geometrical structures of the physical world (e.g., shape, appearance, motion) from raw, casual observations (e.g., Internet photos, videos, audio) in a way akin to human perception. His work received the Best Paper Award at CVPR 2020 and the BMVA Sullivan Doctoral Thesis Prize.