Section: Overall Objectives
Highlights of the year
Excellent results in ImageCLEF evaluation campaigns. LEAR participated in the Photo Annotation and Photo Retrieval tasks of the ImageCLEF 2009 evaluation campaign, which is a part of the Cross Language Evaluation Forum (CLEF). In the first task images have to be annotated automatically with relevant concept names, and in the second relevant images have to be retrieved from a set of 500.000 images given a query image and keywords. For both tasks our results obtained a second place among 19 international participating research teams from industry and academia. CLEF is an activity of the TrebleCLEF Coordination Action under the Seventh Framework Programme of the European Commission. See also http://imageclef.org/2009 .
Action recognition in video. LEAR has recently developed several successful methods for action recognition in video. A method based on bags of spatio-temporal interest points  achieves excellent results in combination with text-based search for retrieving actions  as well as for learning the scene context of actions  . Furthermore, to localize human actions we have developed a human-centric approach. We first extract human tracks and then characterize the actions with 3D histogram-of-gradient descriptors (Sec. 6.4.2 ). This allows to precisely localize human actions in space and time. The interaction with objects (Sec. 6.2.5 ) can further refine the action description.