Anti-Sequences: Event Detection by Frame Stacking

This paper presents a nattiral extension of the new4 introdticed “anti-face” method to event detection, both in the image and in the feattire domains. In the caJe of the image domain (video segtience$ we create spatio-temporal templates by stacking the video framw, and the detection is pe formed on these templates. In order to recogzixe the motion of feattires in a video segtience, the spatial locations of the feattires are modtilated in time, tLxzw creating a one-dimensional vector which represents the event. The followifzg appl t zcaions of anti-sequences are presented: 7) Detection of an oeect tinder 3D rotations in a video segtience simtizated from the CO IL da.tabaJe, 2) VistiaZ speech recognition of spoken word; and 3) Recognition of JJE&OZJ sketched with a laser poifzter. The restiZtizg detection algorhhz is vey fast, and is rebut enotigh to work on small &wags. AZJO, it is capable of dim&hating the desired event-template from a&ray events, and not OE+ events in a ‘hegative traikg set’: