Tuesday, 7th of May

14h30 (room R2014, 660 building) (see location)

Thibault Groueix & Pierre Alain Langlois

(Imagine, ENPC)

Titre: Deep Learning for 3D - Toward surface generation

Abstract

The most striking successes of Convolutional Neural Networks have until
now been demonstrated on images, but the world is tri-dimensional.
Analyzing it should be done directly from the most informative source of
data: 3D content. The development of generative methods for 3D content
would also open up many applications in arts, human-machines
interaction, and education. The first challenge in 3D data analysis is
the choice of a data representation. Indeed, images are naturally
associated to an array (pinhole model), but there is no standard
representation for 3D data. Previous works in deep learning often use
volumetric and point cloud representations. Volumetric representations,
on the one hand, are typically memory intensive and only provide
voxel-scale sampling of the underlying smooth and continuous surface. On
the other hand, point clouds leverage the sparsity of the data and can
provide surface details, but they lack the connectivity between the
points, making it difficult to reconstruct the underlying surface with
high fidelity. Surface based approaches were recently introduced and are
a promising direction.

Short bio

Thibault Groueix has been a PhD student in the Imagine group of Ecole
des Ponts under the supervision of Mathieu Aubry since December 2016. He
is also working in collaboration with Adobe research, supervised by
Mathew Fisher, Bryan Russel et Vova Kim. His PhD work focuses
on synthesizing and analyzing 3D data with Deep Learning.
In particular, the goal is to design novel methods parameterizing 3D
data, and to use these learned parameterizations to reconstruct 3D
content in a friendly format for computer graphics applications. Such
settings include auto-encoding of 3D shapes and single-view reconstruction.

In 2016, he received a Master's degree from MVA. His past academic
experiences include work on rendering in Tamy Boubekeur computer
graphics group (Telecom ParisTech) and work on medical texture
classification in Michael Unser group (EPFL).


Pierre Alain Langlois has been (also) a PhD student in the Imagine group
of Ecole des Ponts under the supervision of Renaud Marlet since November
2017. He is also working in collaboration with Alexandre Boulch at Onera
Palaiseau. His PhD's scope is set on 3D semantic reconstruction of
scenes with a particular focus on indoor/outdoor representations of
buildings. In particular, this subject involves getting interested in
both reconstruction methods (i.e Photogrammetry/LIDAR reconstruction)
and 3D data analysis.

In 2017, he received a Master's degree from MVA. His past related
work includes a project of RGBD-based object recognition (Imagine team +
Pzartech Ltd.), and a multi-object tracking project (Safran Identity &
Security).



Contact: guillaume.charpiat at inria.fr
All TAU seminars: here