Learnable Geometric Reconstruction and UV Parameterization for Digitizing Humans in Loose Garments

Avinash Sharma, IIIT-Hyderabad, India

Friday, 7  October 2022, 15:00-16:00, room 106, Laboratoire Jean Kuntzman, 700 avenue Centrale, Saint-Martin d’Hères

Webex link : https://inria.webex.com/inria/j.php?MTID=m1c3e564d9357c0f8e26f899f631464db

Abstract: This talk primarily aims at providing an overview of our research contribution towards devising learnable paradigms for 3D digitization of human body in loose garments. In the first half, we introduce our novel PeeledHuman representation used for monocular Shape-aware geometric Reconstruction of People in loose clothing (SHARP). We also outline the public release of a high quality 3DHumans dataset collected by our group. Subsequently, we provide an overview of two incremental variants of SHARP exploiting semantic segmentation and face depth prior to obtain monocular reconstruction of the human body in loose garments.  In the second half, we focus on monocular/multi-view garment digitization use-case and present two novel contribution, eXtraction of loose Clothing (xCloth) and learnable object centric UV parametrization with application to textured garment digitization. The xCloth achieves template-free, textured 3D digitization of arbitrary garment styles from a monocular image in a supervised learning setup. Our novel learnable UV parametrization is an object-centric, self-supervised learning framework  to achieve discretization-agnostic surface parameterization of arbitrary 3D objects with both bounded and unbounded surfaces. Finally, if time permits, I intend to also cover 3D terrain authoring/amplification efforts pursued by our research group.

Speaker Bio:  Dr. Avinash Sharma, is a faculty at International Institute of Information Technology Hyderabad (IIIT-H), INDIA, where he is affiliated with the Centre for Visual Information Technology (CVIT). Previously, he worked as Research Scientist at Xerox Research Center India. He graduated with his PhD in Applied Mathematics from INP Grenoble and INRIA Rhone-Alpes, supervised by Prof. Radu Horaud. His current research interests includes 3D computer vision with focus on digitizing humans, large scale 3D asset analysis as well as content generation for AR/VR applications. He has also explored graph based modelling in robotics and currently exploring orthogonal research direction of mathematical modelling of brains by interpreting brain as graph.

 

Comments are closed.