We describe a "bag-of-rectangles" method for representing and recognizing human actions in videos. In this method, each human pose in an action sequence is represented by oriented rectangular patches extracted over the ...