A new way of considering non-rigid structure from motion by considering the possible shapes of a deforming objects as belonging to a non-linear manifold.

CAPTCHA become more and more easy to solve by machines. Why not switching common Turing tests to harder vision problems than OCR ? We present an infrastructure to bridge the gap between computer vision image labeling and stronger Turing tests.

Just a painting stitching project I did for the summer. You can find more info and results here as well as a very cool demo ! (Sorry, it is password protected for Copyright reasons, I hope I can open it soon !)

It is mostly a tweaked version of Brown's excellent IJCV paper

The code is unfortunately copyrighted for now but I can share the help.

The following poster provides a good introduction to our work. LSML is a method for determining a warping from a point on a manifold to its neighbors on the manifold. A direct application of this method is performed on video sequences where the ways of moving on the underlying manifold are learned and then used to move within and out of the training set. The warping is also applied to an unseen frame in order to transfer the tranformations.

When dealing with a crowd of similar objects, severe occlusions happen, leading to a very hard tracking problem. Nonetheless, what if we simply want to count the number of moving objects ? To achieve this goal, we rely on the fact that all the objects in the scene are of the same kind and hence have the same behavior. The main characteristics of our approach are:

• a massively parallel and enhanced KLT tracker
• a new regularization technique for feature trajectories
• simple training data (some video footage with the corresponding number of objects through time)
• a simple learned descriptor, proper to the observed object kind

An art project in which I am doing the computer vision side: SVEN

Briefly, imagine the surveillance cameras are bored with their usual footage and task. What if they wanted to find rockstars in videos for a change ? Well, that is where computer vision (and I) help to create a rockstar detector/tracker. This is basically a BraMBLe tracker with a simple person descriptors (colors, direction, mask, facial expression ...).

The source code for the tracker is here: version 1.0 and the binary here. My coding abilities have improved A LOT so don't judge me on that :). Please email me if you find any bug or way of improving it.

A DVD I participated to. What about asking the leaders in computer vision what they think of the use of their field in movies ?

Computer icons are tiny images conceived to be perceived in a certain way. Therefore a lot less ambiguity exists on the family of higher resolution images that could create the icon (by shrinkage). The trick is to figure out which properties a higher resolution version of an icon should verify.

This implies constraints on region type (gradient, unform...) and perceived colors. Once these are figured out, a cute snake smoothes everything.

The Smart Vivarium is an ongoing project aiming at drastically improving the life of laboratory animals by:

• constantly monitoring them
• keeping statistics
• detecting normal/abnormal behaviors
• minimizing external effects by only relying on video footage