The  Weizmann  Institute  of  Science
                  Faculty of Mathematics and Computer Science
                          Vision and Robotics Seminar
                     Lecture Hall, Room 1, Ziskind Building
                          on Thursday, March 31, 2011
                                 12:00 - 13:00
                                Leonid Karlinsky
                                 will speak on
                           Using body-anchored priors
                    for identifying actions in single images
I will present an approach to the visual recognition of human actions using
only a single image as input. The task is easy for humans but difficult for
current approaches to object recognition, because instances of different
actions may be similar in terms of body pose, and often require detailed
examination of relations between participating objects and body parts in order
to be recognized.  The proposed approach applies a two-stage interpretation
procedure to each training and test image. The first stage produces accurate
detection of the relevant body parts of the actor, forming a prior for the
local evidence needed to be considered for identifying the action. The second
stage extracts features that are anchored to the detected body parts (the
prior), and uses these features and their feature-to-part relations in order to
recognize the action. The body anchored priors we propose apply to a large
range of human actions. These priors allow focusing on the relevant regions and
relations, thereby significantly simplifying the learning process and
increasing recognition performance.
The body parts detection is by itself an interesting, yet difficult and yet not
fully solved task.  Part of the talk will be dedicated to the `chains model'
approach we have developed to address this task.  Detecting an object part
relies on two sources of information - the appearance of the part itself, and
the context supplied by the surrounding parts. The chains model presents a
novel way to use this `internal' context in a flexible manner in order to
propagate information from easy to detect parts (e.g. face) to ambiguous parts
(e.g. hand or elbow) via chains of intermediate image features.
This work was done in collaboration with Michael Dinerstein and Shimon Ullman.
For more details, relevant papers and some example results you are welcome to
(CVPR10, NIPS10).
Technion Math Net-2 (TECHMATH2)
Editor: Michael Cwikel   <> 
Announcement from: Diana Mandelik   <>