Robust Shape Acquisition with Differentiable Rendering
Event details
| Date | 12.07.2021 |
| Hour | 14:00 › 16:00 |
| Speaker | Baptiste Nicolet |
| Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Mark Pauly
thesis advisor: Prof. Wenzel Jakob
co-examiner: Prof. Amir Zamir
Abstract
Differentiable rendering aims at computing derivatives of userspecified
objectives with respect to scene parameters, by differentiating
the light transport simulation, thus allowing to solve challenging
inverse rendering problems. In recent years, considerable progress
has been made towards that goal, allowing to design correct estimators
of derivative quantities.
However, running optimizations with differentiable rendering
remains challenging, especially when it comes to optimizing geometry.
Special care must be taken to avoid bias in the estimated
gradients, and the high sparsity of the gradients obtained make geometry
optimizations tedious. In this report we review three recent
papers that tackle the issue of geometric discontinuities, and review
our work on counteracting the sparsity of the resulting gradients.
Background papers
exam president: Prof. Mark Pauly
thesis advisor: Prof. Wenzel Jakob
co-examiner: Prof. Amir Zamir
Abstract
Differentiable rendering aims at computing derivatives of userspecified
objectives with respect to scene parameters, by differentiating
the light transport simulation, thus allowing to solve challenging
inverse rendering problems. In recent years, considerable progress
has been made towards that goal, allowing to design correct estimators
of derivative quantities.
However, running optimizations with differentiable rendering
remains challenging, especially when it comes to optimizing geometry.
Special care must be taken to avoid bias in the estimated
gradients, and the high sparsity of the gradients obtained make geometry
optimizations tedious. In this report we review three recent
papers that tackle the issue of geometric discontinuities, and review
our work on counteracting the sparsity of the resulting gradients.
Background papers
Practical information
- General public
- Free