Most work on action recognition assumes that the boundaries of the actions in a
test video sequence are given and the labels are unknown. In this talk, we
propose a method to perform joint action recognition and segmentation of test
action sequences. The features extracted from the video sequences are
represented as time series bag of words. Each action class is modeled by an
average exemplar. This average exemplar is learned in a supervised way using
dynamic time warping. The joint recognition and segmentation is done efficiently
by one-pass dynamic programming algorithm. We show results using 2d+t features
extracted from monocular video and scene flow features using stereo.

[ Kaustubh Kulkarni is working in INRIA with Jan Cech, who is the contact person
for this seminar, as I might be absent. Jan Kybic ]