Séminaire du 19/03/2020, 14h00

Aller à : navigation, rechercher

Modeling static and dynamic shape data

In this talk, I will give an overview of my previous work during last 10 years and present my research project in a short term as well as in a longer term. During my ANR project SHARED (2011-2015), I've pioneered the research on feature extraction, correspondence computation, segmentation, and shape retrieval of static and dynamic shapes. The main goal of the project has been to devise shape analysis techniques that exploit the movement data so as to characterize semantically meaningful parts, and guarantee more reliable shape processing than existing techniques that are based on geometric criteria alone. Recently, I have extended some of these works towards the compression of deforming mesh with international colleagues. More recently, I have focused on the deployment of deep learning techniques in the context of 4D human shape modeling. As a start, I chose a human facial modeling as it is an easier topic compared to the body, with more accessibility to a dataset. Unlike most existing methods that treat the expression data as a set of shape instances, I aimed at modeling the temporal evolution of the facial shapes by using RNN (recurrent neural network).

All these efforts have encouraged me to take a bigger step forward: In the framework of an ANR collaborative project (2020-2023), I started to investigate the problem of massive data acquisition, representation, analysis and synthesis of human shape data under motion. Our specific aim is to extend the power of recent deep learning techniques as well as that of 3D/4D modeling, by developing new shape space representations that DNNs can learn over, which can profoundly improve the reconstruction, analysis, and prediction of 3D/4D data, even for the secondary motions and in medical contexts.

In a longer term, I will continue to investigate applications of my research work in medical context and establish collaboration with other subareas. Both my previous work (feature extraction, surface registration, etc.) and earlier achievements (shape space construction, statistics based shape estimation, etc.) seem to be scientifically relevant in finding medically significant applications. This line of research has been already initiated with women’s breast data, in my recent collaborative work with radiation oncologists at Centre Paul Strauss. I continue discussions and exchange with researchers at ICube who are currently active in this direction.