Unsupervised Object Discovery from static camera is at stake in this PhD offer. Such a method is particularly useful for automatic annotation of videos as it avoids an extremely time consuming work for person when it is done by hand. It is also very interesting for forensic application or for learning contextualizing object detectors that can be in various domain of application (real time object detection or tracking, etc). Object discovery in the scene will be done in an unsupervised way by modeling and clustering recurrent visual observations. A special attention on the choice and constitution of visual features to analyze will be done, according to recent works on “Bag of Word” and “Unsupervised Feature Learning”. The learning method for class discovery will be another key step that will be studied with care with method such as “Topic modeling”. In a second phase, use of contextualized detectors on the studied scene will be studied in order to refine video annotation. These contextualized detectors will be trained using examples drawn from the classes previously discovered.

This position is open until it is filled.

Département: Département Intelligence Ambiante et Systèmes Interactifs (LIST)
Laboratory: Vision & Ingénierie des Contenus (SAC)
Start Date: 01-10-2015
ECA Code: SL-DRT-15-0421
Contact: bertrand.luvison<στο>cea.fr