Visual Object Recognition: Can A Single Mechanism Suffice?

-1-Visual Object Recognition: Can A Single Mechanism Suffice?Michael J. TarrBrown University“In actual, physical life I can turn as simply and swiftly as anyone. Butmentally, with my eyes closed and my body immobile, I am unable toswitch from one direction to the other. Some swivel cell in my brain doesnot work. I can cheat, of course, by setting aside the mental snapshot ofone vista and leisurely selecting the opposite view for my walk back tomy starting point. But if I do not cheat, some kind of atrocious obstacle,which would drive me mad if I persevered, prevents me from imaginingthe twist which transforms one direction into another, directly opposite. Iam crushed, I am carrying the whole world on my back in the process oftrying to visualize my turning around and making myself see in terms of“right” what I saw in terms of “left” and vice versa.”—Vladimir NabokovHow do humans recognize 3D objects? This simple question leads to surprisinglycomplex answers. Indeed, object recognition is sufficiently difficult that state-of-the-artcomputer vision systems can only perform the most rudimentary visual tasks and, eventhen, only under highly constrained conditions. At the heart of what makes visualrecognition difficult are two factors. First, we live in a world made up of 3D objects, yetonly receive 2D stimulation on our retinae as sense input. Second, we live in a highlyvariable world in which images of objects change constantly due to transformations insize, position, orientation, pose, color, lighting, and configuration. The challenge is toderive a consistent mapping from a potentially infinite set of images to a relativelysmall number of known objects and categories. It is a problem that the human visualsystem routinely and effortlessly solves.How the mammalian brain solves the problem of visual recognition has been a topic ofstudy since the early days of cognitive science. David Hubel and Torsten Wiesel(1959) received the Nobel Prize for their discovery of organized columns oforientation-tuned neurons in cat visual cortex. This critical result appeared to capturean important facet of visual processing—a visual system that is sensitive to edges(boundaries between regions of light and dark) positioned at different orientations inspace. Once the particular orientations of edges are known, it seemed only a smallstep to “connect the dots”—joining edges into more complex descriptions of objectshape. Edge-based representations appeared ideal for recognition: shape definingedges often capture the critical features of objects and remain relatively invariant overmany image transformations. Thus, most vision scientists came to believe that the goalof vision was to derive or reconstruct an edge-based description of object shape.It was this belief that drove David Marr to develop two ideas that have dramaticallyinfluenced the study of visual object recognition. The first was a computational theoryin which Marr and Ellen Hildreth (1980) proposed what they saw as the processingconstraints needed to build a successful edge detector. They observed that animplemented version of their detector behaved much like some subclasses of visualneurons (so-called “simple cells”). Once Marr had a plausible algorithm for finding