A Shift in Vision

A Shift in Vision: Recognition, Context and Depiction
by Ron Gallagher
Abstract: Recent work on the human visual system and on the development of computer based visual recognition systems indicates that shape-matching is an unworkable basis for human or machine recognition of objects, scenes and pictures. Studies of ‘gist views’ have shown that scene recognition is prior to object recognition. Therefore context, broadly construed, has a primary role in object recognition. This paper argues that we recognise objects in pictures using the same mechanisms that we use to recognise objects in the real world. It therefore follows that context, as opposed to shape matching, is the key to understanding how we see content in depictions. Vision researchers are finding it difficult to shake off a theoretical model of the human visual system which characterizes it as a shape matching system. The predominance of the shape-matching model has focussed on the problems which the human visual system must overcome in interpreting the two-dimensional geometry of light stimulus on the retina as three-dimensional objects. For example, V.S. Ramachandran begins his 1988 paper ‘Perceiving Shape from Shading’ thus: Our visual experience of the world is based on two-dimensional images: flat patterns of varying light intensity and color falling on a single plane of cells in the retina. Yet we come to perceive solidity and depth. We can do this because a number of cues about depth are available in the retinal image: shading, perspective, occlusion of one object by another and stereoscopic disparity. In some mysterious way the brain is able to exploit these cues to recover the three-dimensional shapes of objects (Ramachandran, 1988, pp. 76-83). This unfortunate emphasis on the two-dimensionality of the retinal image has led vision researchers and philosophers to speculate that the mechanisms whereby the human visual system recognises objects in pictures echoes the reconstruction of...