Jérôme - Consultant C++

Work Experiences
October
2018–Now
CNRS, Sophia Antipolis, Research Engineer.
Working on SLAM and deformable object SLAM. With link towards semantic annotations
based on Artificial Intelligence. Algorithms resort to a 3D mesh with computation on graphics
card (OpenGL and GLSL).
February
2017–
September
2018
Freelance, Paris, Computer vision developer.
I have been working on Neural Network acceleration regarding sorting algorithms and Non
Maximal Supression acceleration based on pre-computed grid regions within CUDA
Janauary
2016–
February
2017
Navya, Paris, Computer vision developer.
I am developing computer vision algorithms in the field of stereovision and machine learning
for autonomous vehicles. At the present time I have developed real time algorithms using
CUDA and I am succesfully working on a fast version of the adaptative support weight (ASW)
stereovision approach. The complexity of the ASW is W × H. W and H being the size of the
window search. My approximation leads to the same quality with a complexity log(W).
September
2015–
January
2016
Abbeal, Paris, Computer vision developer.
I have developed an image based calibration algorithm for the company Aldebaran. The goal
is to give to the robot pepper the ability to capture the 3D structure of a scene in order to
perform classification, detection and tracking.
September
2014–now
Bertin Technologies, Paris, Computer vision developer.
I have been developing a face recognition algorithm. The objective is to reconstruct the 3D
face from a single photograph, then it is possible to frontalize the face and synthetize a new
picture of the face from a frontal point of view. The last step consists in a face recognition
algorithm from this novel picture. I resort to neural network and stereovision approaches for
this project.
January
2012–
September
2014
Telecom SudParis, Paris, 3D Research.
Conception of a face tracker algorithm based on image and depth map, for HD and real time
application. The objective consists in fitting a morphable model of the face on to a 3D target
and the image using SVM. The algorithm within a variational framework can be applied for
movie industry and video game industry. One part for movie industry is using dense algorithms
with OpenGl, and the part devoted to video games is using the sparse version with CUDA
programming.
Development of several Autodesk Maya plugins. The project has been published to GTC
Nvidia. The following video shows highly accurate results obtained with a synthetic dataset:
jerome.********.********/ImageDepthTracker.avi

April
2011–2012
Vidatis, Paris, Computer vision developer.
I develop a device and program with OpenCV which consists in detecting in real time defaults
of serigraph on bottles of perfume. In a first step we learn the bottle serigraph by capturing
the bottle with different viewpoints. We get reference images. Then we retrieve contours and
transforms between images using three main ingredients: Delaunay triangulation of the image,
image rectification, and optical flow without vertical motion. Once we have serigraph contours
and image transforms, we retrieve accurately the 3D serigraph and camera parameters.
Later we analyse the serigraph of a candidate bottle within one second. First we capture
three candidate images with three different viewpoints. We calibrate these cameras reproject
candidate images on reference frames with OpenGL. Then we segment images and detect
robustly serigraph errors using redundancies between the three candidate images.
2007–2011 Ecole des Ponts / LIGM-IMAGINE, Marne-La-Vallée, Ph.D Thesis on SpatioTemporal Reconstruction and Structure from motion.
The first part of my Ph.D was devoted to spatio-temporal reconstruction of dynamic scene
captured from synchronized cameras. This method allows to get a dense and accurate
reconstruction, and also to track points over time. This method consists in minimizing an
energy integrated on the image pixel grid. This program was developped with CGAL to deal
with mesh, and with OpenGl and Cg to compute efficiently the gradient under graphics card.
The second part of my Ph.D is devoted to structure from motion. I proposed a novel method
allowing to retrieve the camera parameters (rotation, translation, focal length) without any prior
knowledge. This approach allows to deal with cyclicity constraints and loopy configurations on
the graph of cameras. As shown on IJCV paper it is possible to turn polynomial loop constraints
into linear constraints. Finally once can calibrate a camera network using simple linear
programming. The proposed framework benefits from the quickness of linear programming
and cancels drift for loops made of numerous images.
2008–2010 Ecole des Ponts ParisTech, Marne-La-Vallée, Teaching, C++ Programming courses.
2007 Ecole Normale Supérieure, Paris, Research internship, 5 month.
Development of a sequential structure from motion software based on successive bundle
adjustments. The objective is to retrieve cameras positions and focal lengths. To achieve
this goal I used an incremental procedure described on the book “Multiple View Geometry in
Computer Vision” from R. Hartley and A. Zisserman.
Supervisors: Jean Ponce, Andrew Zisserman.
2002–2006 Technip, Paris, System management developer.
Conception along with intervening people, and development of web-based management system
to deal with documents’ workflow. Mixed client and web applications with management of
roles for intervening people during the document’s life. Programming of agents updating
documents on the server and between the replicated databases all over the world.

Selected Talks
03/2014 High-Resolution Facial Performance Capture Using CUDA. NVIDIA GTC 2014. San
Jose. USA.
02/2011 Spatio-temporal reconstruction and structure from motion with Loop constraints
linearization. TOSHIBA Cambridge.
05/2009 A global camera network calibration method with Linear Programming. 3DPVT 2010.
Paris. France.