Deep learning under hard constraints for 3D shape representation
Event details
| Date | 12.07.2021 |
| Hour | 10:00 › 12:00 |
| Speaker | Nicolas Talabot |
| Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Wenzel Jakob
thesis advisor: Prof. Pascal Fua
co-examiner: Prof. Nicolas Flammarion
Abstract
Representation and generation of 3D shapes has
become a significant line of research in the computer vision
community. The real objects that are thus represented may
have a set of constraints, either physical or technical, that
should be satisfied. As deep learning is often used to create
these representations, our research focuses on imposing hard
constraints over their output. We are currently working on
developing training algorithm based on the Lagrangian while
comparing to soft constraints. In this report, we present three
existing works relating to these problems. We first discuss the
introduction of a differential constrained optimization algorithm
for neural networks, then an newer work proposing a training
based on a primal-dual formulation of the Lagrangian, and
eventually take a look at a model representing 3D shapes as
a set of separate parts respecting some structural constraints.
Background papers
exam president: Prof. Wenzel Jakob
thesis advisor: Prof. Pascal Fua
co-examiner: Prof. Nicolas Flammarion
Abstract
Representation and generation of 3D shapes has
become a significant line of research in the computer vision
community. The real objects that are thus represented may
have a set of constraints, either physical or technical, that
should be satisfied. As deep learning is often used to create
these representations, our research focuses on imposing hard
constraints over their output. We are currently working on
developing training algorithm based on the Lagrangian while
comparing to soft constraints. In this report, we present three
existing works relating to these problems. We first discuss the
introduction of a differential constrained optimization algorithm
for neural networks, then an newer work proposing a training
based on a primal-dual formulation of the Lagrangian, and
eventually take a look at a model representing 3D shapes as
a set of separate parts respecting some structural constraints.
Background papers
- Constrained differential optimization - weblink (pdf)
- A Primal-Dual Formulation for Deep Learning with Constraints - weblink/pdf
- SDM-NET: deep generative network for structured deformable mesh - weblink (pdf)
Practical information
- General public
- Free