Recent advances in deep learning have enabled new generative models for 3D objects, in particular manufactured shapes (e.g. chairs, tables, airplanes, cars, ...). The most state-of-the-art methods in this space are "structure-aware": they take into account the parts of which the object is made and how those parts are assembled. However, there has been relatively little progress in generating 3D characters: people, animals, and other such 'living agents.' Generative models of characters would be useful for content creation in animation, visual effects, and games; they would also find application in 3D reconstruction applications (for example, marker-less tracking and reconstruction of animals). Character modeling presents unique challenges not shared by manufactured shapes: characters can vary widely in body structure, shape, and pose, and they typically do not decompose into assemblies of parts (since they are mostly organic). How can we design structure-aware generative models of 3D characters and learn them from data? This project seeks to answer this question. Our hypothesis is that a good representation for a character's structure is its skeleton.
Zhan, Xiao,
"Learning Structure-Aware Generative Models of 3D Characters"
(2021).
Summer Research Symposium.
Brown Digital Repository. Brown University Library.
https://doi.org/10.26300/9tz8-v671
Each year, Brown University showcases the research of its undergraduates at the Summer Research Symposium. More than half of the student-researchers are UTRA recipients, while others receive funding from a variety of Brown-administered and national programs and fellowships and go …