Hierarrhicd Wfmsolution Data St" and Algorithm for Foveal Vision System
César Bandera, Peter Dale Scott · 1990
There are two ways by which resolution may be dispersed over a scene which is to be studied. The spahd resolution of the imager ma vary over its field of view, yielding images with finer delail in some image re ions &an others. or the field of view of the imager may be allocated temporally, yiefding longer gaze times and shorter revisit times to me regions of the scene than others. For static Scenes in the presence of noise and sensor imperfections, the achieved resolution at any point in the scene improves monotonically with each of these factors. other scenes, the fo? of dependence can vary but both remain significant determinants of achieved resoluhon. Both space-variant and time-variant resolution resource allocation are an integral part of essentially every biological vision system. Recording of e e movements while an individual is studymg an unfamiliar pamting [I] reveals smalffocj of auention which are revlsited frequently and held for longer-lhan-average durahons (intcreslmgly, different individuals select different regions to favor). Furthermore, each snapshot is ima ed by a retinal sensor which is of much higher spatial resolution in a small cendre ion the fovea, lhan elsewhere In the human eye resoluhon rolls off smoothlyf,the central fovea to the riphery of the field of Aew with a decline of angular resoluhon of se vera^ orders opemagnitude. Other creatures possess fovea of stnking diversity of shape and rolloff, but all have one or more such high resolution regions surrounded by zones of graded resolution reduction. While the evolutionary process has@ almost unanimously to implement resolution resource allocation by both space-varian~ and time-variant mechanisms in its vision systems, the machine vision community has voted almost unanimously to the contrary. There are several ood reasonsfor this basic disagreemenl. Of prim importance,evolution has hajsome th~ billion years head start on technology. W% respect to the time-vanant mechanism, significant efforts have recently been ma& to catch up. Gaze conuol b feedback requires sophisticaled and expensive mechanical servo loo wxose dynamics must be based on unknown strategies for deciding where to IooE&xt and for how long. The study of such active rception systems and strategies. pioneered in the laboratories of Bajcsy 121, Burt E], Ballard [ 11 and Aloimonos 141. while of relatively recent origin, is achieving recognition as an impt and continuing line of investigation in machine vision and machine percepuon. With respect to the space-variant mechanism, less progress has been made. While devices incorporating time-variant resolution allgation, i.e. perception, may be usefully studied in the absence of space-vanant or vision, ima ers incorporaung foveal vision are of little use without mechanisms for sleering taem intelligenll and accurately across a scene. The foveal vision problem (!.e. the problem ordesigning data structures and algorithms to make effechve use of unages ipal data srm~ture proposed is the foldmg of fovea! data mto