Selected Publications

We present a method to extract per-point semantic class labels from aerial photogrammetry data. Unlike previous point cloud classification methods that rely exclusively on geometric features, we incorporate color information yielding a significant increase in accuracy. We test our classification method on three real-world photogrammetry datasets generated with Pix4Dmapper Pro, achieving good accuracy and generalization.
In ISPRS, 2017

Local Binary Descriptors (LBDs) have become very popular for image matching tasks, especially when going mobile. However, their ability to carry enough information to infer the original image is seldom addressed. We leverage an inverse problem approach to directly reconstruct the image content from LBDs and without requiring either a prior learning database or non-binarized features. Interestingly, our reconstruction scheme reveals differences in the way different LBDs capture and encode image information.
In TPAMI, 2013

Local Binary Descriptors (LBDs) are good at matching image parts, but how much information is actually carried? Surprisingly, this question is usually ignored and replaced by a comparison of matching performances. Here, we directly address it by trying to reconstruct plausible images from different LBDs such as BRIEF and FREAK and show that this task is achievable using only the information in the descriptors.
In ICPR, 2012

Vision is a natural tool for human-computer interaction, since it provides visual feedback to the user. It requires however the fast and robust computation of motion primitives. We apply here some recent mathematical results about convex optimization to the TV-L1 optical flow problem. At the cost of a small smoothing of the Total Variation (TV), the convergence speed of the numerical scheme is improved, and realtime performance is achieved using the OpenCL framework.
In ICIP, 2009

Recent Publications

. From bits to images: Inversion of local binary descriptors. In TPAMI, 2013.

Preprint PDF Code Project Project

. Fast TV-L 1 optical flow for interactivity. In ICIP, 2009.

Preprint PDF Project

Recent Posts

More Posts

When starting with Deep Learning on your own (without any legacy code or compatibility constraint), it may be daunting to choose one among the many frameworks available.


[Update 22.03.2018: link to correct Youtube publication.]

U-Net was proposed in 2015 for medical image segmentation. You can find the original paper, along with some video introduction on the project homepage. Its structure is relatively simple and shallow, so it seems to be well fitted for a learning work.


It turns out I completed my PhD in 2012 just before Deep Learning started to boom and be the Next Big Thing in Computer Vision. The various pieces were already there (neural networks, back propagation, large scale databases, GPGPU…), but still, I somehow managed to slip through it. Later at work, I’ve used standard Pattern Recognition approaches for realtime applications, and later the more complex and efficient Gradient Boosted Trees, but still, no luck at trying Deep Learning.


I’m restarting this blog after keeping it quiet for a few years. What changes to expect? Well, quite a bit: I’m back to Computer Vision, playing with photogrammetry pipelines, meshes, classifiers… I’m moving into product management this blog is now published thanks to hugo with the Academic theme the comments are disabled, but you can interact through the contact address or directly via social networks. I’ve been lucky (and grateful) to attend new events and meet new communities last year, primarily at ISPRS (Hannover) and the Product Management Festival in Zürich.


Story of a bug that I introduced in my codebase while making it parallel.



Image reconstruction from Local Binary Descriptors

Inferring image content from local descriptors.


Continuing education

  • EUROSAE: MTS-014 Introduction to image manipulation. I am teaching 3 days (out of the 6) of this course directed by Bertrand Collin.

Academic courses

While a PhD student at EPFL’s Signal Processing Lab 2, I’ve been a TA for the following courses:

  • Introduction to Analog Signal Processing (SV and MT bachelors, duration 1 semester)
  • Introduction to Digital Signal Processing (SV and MT bachelors, duration 1 semester)
  • Advanced Signal Processing (EE master, duration 1 semester)

While in France, I’ve also given C/C++ programming labs for ENSTA-ParisTech new students.