|
Ph.D Defense J. Kooij - Talks by B. Leibe (RWTH Aachen) and J.M. Odobez (IDIAP) on 04.12, Aula Univ. Amsterdam
The Ph.D. Defense of Julian F. P. Kooij titled “Generative Models for Pedestrian Track Analysis” will be held Friday, 4 December at 11:00, at the Aula of the Univ. of Amsterdam (Singel 411, Amsterdam). Prior to this, from 9:30-10:30 at the same location, there will be two interesting talks on computer vision and machine learning by prof. Bastian Leibe (RWTH Aachen) and dr. Jean-Marc Odobez (IDIAP). For titles and abstracts, see below. You are all cordially invited to attend!
---------
Shape Cues for Generic Object Tracking in Street Scenes
Prof. dr. Bastian Leibe, RWTH Aachen
http://www.vision.rwth-aachen.de/people/leibe
Most vision based systems for object tracking in urban environments focus on a limited number of important object categories such as cars or pedestrians, for which powerful detectors are available. However, practical driving scenarios contain many additional objects of interest, for which suitable detectors either do not yet exist or would be cumbersome to obtain. In this talk, I will present recent research on leveraging shape cues for generic object tracking that can complement the traditional detectors. At the core of this approach is an object-centric shape representation that makes it possible to robustly track unknown objects while reconstructing their 3D shape, to perform shape-based multi-category classification, and to support learning of person-object interactions.
---------
Exploiting long-term connectivity and visual motion in CRF-based multi-person tracking
Dr. Jean-Marc Odobez, IDIAP
http://www.idiap.ch/~odobez/
In this presentation, i will introduce a conditional random field approach to solve the tracking-by-detection task in which we model pairwise factors linking pairs of detections and their hidden labels, as well as higher order potentials defined in terms of label costs. To the contrary of previous approaches, our method has important characteristics that make it successful: (i) long-term connectivity between pairs of detections; (ii) exploitation of models similarities as well as dissimilarities between detections, based on position, color, visual motion cues; (iii) time-interval sensitive association scores; (iv) feature-specific confidence scores, which aim at weighting feature contributions according to their reliability; (v) two stage unsupervised learning of pairwise potential parameters. Experiments on PETS’09, TUD, CAVIAR, Parking Lot, and Town Center public data sets show the validity of our approach, and similar or better performance than recent state-of-the-art algorithms.
|
|
