BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Deep learning under hard constraints for 3D shape representation
DTSTART:20210712T100000
DTEND:20210712T120000
DTSTAMP:20260504T193323Z
UID:f723149d5a2ecff06f043ac87c8bd6bcd12c462f654b3a34000b864a
CATEGORIES:Conferences - Seminars
DESCRIPTION:Nicolas Talabot\nEDIC candidacy exam\nexam president: Prof. We
 nzel Jakob\nthesis advisor: Prof. Pascal Fua\nco-examiner: Prof. Nicolas F
 lammarion\n\nAbstract\nRepresentation and generation of 3D shapes has\nbec
 ome a significant line of research in the computer vision\ncommunity. The 
 real objects that are thus represented may\nhave a set of constraints\, ei
 ther physical or technical\, that\nshould be satisfied. As deep learning i
 s often used to create\nthese representations\, our research focuses on im
 posing hard\nconstraints over their output. We are currently working on\nd
 eveloping training algorithm based on the Lagrangian while\ncomparing to s
 oft constraints. In this report\, we present three\nexisting works relatin
 g to these problems. We first discuss the\nintroduction of a differential 
 constrained optimization algorithm\nfor neural networks\, then an newer wo
 rk proposing a training\nbased on a primal-dual formulation of the Lagrang
 ian\, and\neventually take a look at a model representing 3D shapes as\na 
 set of separate parts respecting some structural constraints.\n\nBackgroun
 d papers\n\n	Constrained differential optimization - weblink (pdf)\n	A 
 Primal-Dual Formulation for Deep Learning with Constraints - weblink/pdf
 \n	SDM-NET: deep generative network for structured deformable mesh - webl
 ink (pdf)\n\n\n 
LOCATION:
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR