Poster at ICCV workshop

My good friends Nicolas Thome and Matthieu Cord presented a joint work on kernel learning for computer vision at the 1st IEEE Workshop on Kernels and Distances for Computer Vision.

We worked on nice algorithms for combining different visual descriptors in the kernel framework and this poster is a kind of preview for some not yet published works.

link to the pdf

Do more backups

Last week my laptop's hdd crashed, probably because of the high temperature we do have now in France (!!!). Fortunately enough, I was able to recover all my emails, last codes and writings. Unfortunately though, I lost quiet a lot of experimental data. That's life.

I'll definitely do more backups. Right now, I'm using rsync in the old fashion:

rsync -av --delete Archives login@backuphost:

P-H. Gosselin gave me a nice trick to deal we this. You put all things important in only one folder 'Archives', like your '.thunderbird' folder for instance. Then you symlink them back to where they're usefull (e.g. ln -s ~/Archives/.thunderbid ~/.thunderbird). This way, you only need to save the Archive folder, which simplifies the backup script.

To automate things, I added some fancy gui using zenity:

if zenity --question --text="Do the backup?"
then rsync -av --delete ~/Archives login@backuphost: | zenity --progress --pulsate

I schedule the script every day around lunch using gnome-schedule. Thus, I have a nice popup which asks me to allow the backup to proceed. One thing that would be nicer, would have to know the output of rsync, which at the time is eaten by zenity.

Paper Accepted in ICIP'2011

We (Me and P-H Gosselin) have an oral presentation at ICIP'2011 in Brussels. It's about a new framework combining kernels on bags and vector representations for improving image similarities. The results a very promising both in near-duplicate search and in image classification. As soon as the camera ready paper is done, I'll upload it here.

Seminar in Brazil

I will be in Brazil for the next two weeks. I've been invited for a talk at UFMG scheduled on Friday 03/12/21010 morning at UFMG, Belo Horizonte:

Content based multimedia retrieval and indexing

In multimedia content based retrieval, we usually tackle three different user aims when using the system: the search for a target document, the interactive browsing of a multimedia database, and the categorization of an arbitrary number of documents into predefined classes. This talk will provide an overview of my work in these three areas. The first part will be on the retrieval of urban scene photographs (as part of the french iTowns project). The second part will be on active statistical learning tools for interactive browsing, especially in the context of browsing distributed databases and collaborative systems. The last part will be on image categorization, where I will present machine learning techniques for efficiently combining multiple description channels such as shape, color and texture.

Special Session in CBMI 2011

Our special session (with P.-H. Gosselin) for CBMI 2011 about "Collaborative Learning for Interactive and Long Term Retrieval in Multimedia Databases" has been accepted!


Machine learning techniques are now widely used in Content-Based Multimedia Indexing (CBMI) systems to bridge the semantic gap. Most of these methods are designed around a single user, who is invited to build a query made of image annotations. Depending on the difficulty of the searched concept, the user gives more and more annotations, until he is satisfied by the results. In such a scenario, better retrieval systems are those which return better results with less annotations.

Then, once such systems are deployed, many users perform their searches independently. However, it is common that several users are looking for similar visual concepts. That means that a lot of knowledge can be shared among users. In other words, if previous users already found several concepts, a smart retrieval system should take advantage of this previous knowledge in order to speed up the next retrieval sessions.

This problematic is poorly addressed by the content based multimedia retrieval community. The aim of this special session is to first present the latest methods in this scope, but also to invite new researchers to address these learning problems.

Scope and topics

The scope of this session is to cover innovative indexing systems which break the single user paradigm. Such systems are facing many learning problems. In that case, the main question is how to efficiently combine these learning tasks to improve the results. A first example would be the collaborative learning of a single concept and a second one the long-term learning of unknown concepts.

The special session seeks for papers describing original work in the following areas:
  • multi-users retrieval systems
  • collaborative learning of a single concept
  • long-term learning of unknown concepts
  • fusion of retrieval sessions
  • concepts discovery
  • active learning in collaborative learning context


  • Philippe Henri Gosselin, ETIS Lab - ENSEA, France
  • David Picard, ETIS Lab - ENSEA, France

«previouspage 6 of 6