Coordinate Transformations in Object Recognition

摘要

A basic problem of visual perception is how we recognize objects after spatial transformations. Three central classes of findings have to be accounted for: (1) Recognition performance varies systematically with orientation, size, and position. (2) Recognition latencies are sequentially additive, suggesting analog transformation processes. (3) Orientation and size congruency effects indicate that recognition involves the adjustment of a perceptual coordinate system. While existing models of object recognition are unable to account for all three types of findings, the data can be explained by a transformational model of recognition (TMR), which relies on coordinate transformations and multiple representations. Recognition is achieved by transforming a generic perceptual coordinate system that defines the correspondence between positions specified in memory and positions in the current visual field so that memory and input representations are aligned. TMR is an analog model of recognition, based on analog (image -like) representations and analog transformation processes. TMR is compatible with models in computational neuroscience proposing that object recognition involves coordinate transformations, implemented by neural gain (amplitude) modulation. These gain modulation processes correspond to transformations of receptive fields in upstream cortical areas. Coordinate transformations seem to underlie both object recognition and visuomotor control, and may be regarded as a general processing principle of the visual cortex. TMR discriminates between compensation processes in object recognition and mental imagery; coordinate transformations in recognition are more closely related to visuomotor control than to mental rotation. This framework has several advantages and overcomes arguments that were raised against alignment models of recognition.