AI Engineer


Artificial Inteligence Engineer from Polytech’Nice-Sophia’s university graduated in Electronics in 2015. I’m currently working at the CNRS as a research Engineer in computer vision in the Sparks team. Member of the i3s laboratory, my work is focused on automatic 3D point clouds scenes understanding.


Point Cloud DPM (Deformable Part Model) for 3D scene object detections
Extension of the Deformable Part Model for 3D point clouds, based on a linear Structural Latent SVM. We obtains confident results during the test of the algorithm on easy scenes. Unfortunately, the current algorithm is robust to scale and translation transformation but not to rotation. Our original goal was to fix this defect but our approach seems to failed because of the complexity of the problem in 3D. Considering the power of DNN with complex issues, the existing methods which allows spatial transformation invariance and the fact that our method can be transposed in DNN, we started to translate our approach into a DNN. The work is still in progress, because of time constraints, but we expect that this step will allow us to solve the rotation robustness issue and even improve the detection score of our original architecture.

3D Shape Retrieval Engine for archeological artefacts
Based on the association of similarity search technique and active learning, this algorithm is designed to help non-expert users to make their way in the machine learning territory. Our Content-Based 3D shape Retrieval (CB3DR) solution aims at scanning an object on the fly with a low-cost 3D sensor and retrieving similar shapes or informations of similar objects from a database using the 3D point cloud acquired.

Gender and Age Estimation of Pelvis bones using Deep Learning
Archeological tool develped for estimate gender and age of a human pelvis bone acquired from a low cost 3D sensor. We decided to treat those problems distinctly for practical reasons. The gender can be easily estimated throught shape analysis of the bone. This was made using the PointNet network. On the other hand, the age is harder to retrieve because of the quality of the sensors we used so we used the DeepTEN network which was our best option at that time.

KPPF: Keypoint-Based Point-Pair-Feature for Scalable Automatic Global Registration of Large RGB-D Scans
One of the most important challenges in the field of 3D data processing is to be able to reconstruct a complete 3D scene with a high accuracy from several captures. Usually this process is achieved through two main phases: a coarse, or rough, alignment step then a fine alignment. In this article we propose an automatic scalable global registration method (i.e. without arbitrary pose of the sensor) under the following constraints: markerless, very large scale data (several, potentially many millions of points per scans), little overlap between scans, for more than two or three dozens of scans, without a priori knowledge on the 6 degrees of freedom. Here we only address the coarse alignment, and consider the fine alignment step tackled by dedicated existing approaches such as Iterative Closest Point (ICP) [3]. We evaluate thoroughly our method on our own dataset of 33 real large scale scans of an indoor building. The data presents some pairs of scans with very little overlap, architectural challenges (a patio and a rotunda open through several levels of the buildings, etc), several millions of points per scan. We will make this dataset public as part of a benchmark available for the community. We have thus evaluated the accuracy of our method, the scalability to the initial amount of points and the robustness to occlusions, little scan overlap and architectural challenges.