I am co-advised by Iasonas Kokkinos and Niloy J. Mitra. My research interests are centered around Deep Learning, Computer Vision and Computer Graphics. The focus of my PhD is on Deep Learning for 3D Point Clouds.
From 2017 to 2018, I worked at Société Générale, as part of the Research Group within the General Inspection. I worked on a wide range of projects there in both Natural Language Processing and in Computer Vision.
I hold two MSc, one in Applied Mathematics from CentraleSupelec, where I graduated with the Highest Honour with the best overall grade of my class and a second one in Machine Learning and Computer Vision (Master MVA) from Ecole Normale Supérieure Paris-Saclay (formerly called ENS Cachan).
Eric-Tuan Lê, Iasonas Kokkinos, Niloy J. Mitra
In this work, we train deeper and more accurate point processing networks by introducing three modular point processing blocks that improve memory consumption and accuracy: a convolution-type block for point sets that blends neighborhood information in a memory-efficient manner; a crosslink block that efficiently shares information across low- and high-resolution processing branches; and a multi-resolution point cloud processing block for faster diffusion of information. By combining these blocks, we design wider and deeper point-based architectures.
We report systematic accuracy and memory consumption improvements on multiple publicly available segmentation tasks by using our generic modules as drop-in replacements for the blocks of multiple architectures (PointNet++, DGCNN, SpiderNet, PointCNN).
Academic Research Projects
Spectral estimation of Tip-Timing signals, in partnership with Safran
This work, in partnership with Safran, aims at making quality control on motor rotors in operations easier. Tip timing analysis is one of the main method for non-destructive control on rotors. However, the recorded signal is sub-sampled and non-uniformly sampled leading to aliasing. In this work, we designed a set of methods for tip-timing spectral analysis for real world signals. We proposed to solve the problem by two different formulations: (i) in the form of a LASSO approach and (ii) in the form of a Multiple Measurement Vector approach.
We showed that our approaches are robust to a SNR as low as −7 dB.
Detection of Copy Move Forgery on real world images [Project] [Code Poisson Editing] [Code SIFT] [Code PatchMatch]
Copy-move forgery is a forgery method that is non trivial to detect because the statistics of the images are kept intact. It is thus usually used to hide or alter visual information. In this work, we built two pipelines to automatically detect copy-move forgery within an image.
We generated fake images by copy-move forgery using Poisson editing that relies on the idea that human vision is usually more sensitive to the Laplacian. The generated images have been used to test two pipelines: one relying on dense matching with PatchMatch and the other on sparse matching with the SIFT approach.
Edge and pattern detection on images, in partnership with Microvision
In this work, in partnership with Microvision, I designed a pattern detection pipeline to automate visual control. The pipeline automatically extracts the contours from an image and then detects straight lines and circles within an images via the Hough Transform.
We tested our pattern detection pipeline to automatically analyze blood tests by detecting and measuring the diameters of radial immunodiffusion circles.