Speaker: Sébastien Poullot
Date: July 28th, 2014
Place: room 214, Faculty of Science Bldg. 7, Hongo Campus, The University of Tokyo
In the latest years a new family of kernels for visual contents search or classification has raised: the Fisher vector-like descriptors, which can be seen as a set of features match kernel. Such a description embeds a set of local features in a global one, this is not new, an historic well known representation is the bag of feature approach. However a recent generalization of this approach broke through the discretization paradigm adopted by the community for dozens of years (for better generalization and faster
computation) and opened the door of continous descriptions of visual contents. Furthermore such description can embed a complex distance function usually used between set of points in order to express probabilistic concepts.
The presentation start with a bit of history to introduce, to mention but a few, the match kernels or the SVM, then, following a concrete example, an application to video event matching, it will sidetrack on major concepts, in order to show the very strong potential of the set of features match kernel.