Dorsal–ventral integration in object recognition

Abstract

The idea of two parallel hierarchical pathways in vision has fueled a great deal of research and enhanced our understanding of visual processing in the brain. However, after 25 years, it has become clear that the earlier distinctions in terms of neuroanatomy and functional dissociation are less pure than originally considered. Dorsal visual areas may exhibit object-selective responses and many 3-D cues of shape, particularly structure-from-motion, appear to be computed exclusively by dorsal areas. These findings imply a more important role for dorsal visual areas in object recognition than previously considered and also place restrictions on the nature of ventral object representations. These representations will need to include information about the objects in 3-D, making them more viewpoint-invariant. They will also need to be invariant to the 3-D cue used to describe them. Through the discussion of relevant findings in psychophysics, single-unit electrophysiology, neuroanatomy and functional imaging, I suggest that these qualities are indeed present in ventral stream representations. Thus dorsal visual areas that extract 3-D structure of shapes from certain cues, can relate these representations to cue-invariant and view-invariant representations in the ventral stream.

Introduction

In 1982, an idea was presented that dramatically influenced thinking about the primate visual system. Ungerleider and Mishkin (1982), based on the pattern of behaviour following lesions to dorsal (occipito-parietal) and ventral (occipito-temporal) regions of the monkey cortex, suggested that the visual cortex can be decomposed into two pathways—a dorsal pathway concerned with spatial properties of vision (answering the question “where?”) and the ventral pathway concerned with identification of the visual objects (answer the question “what?”). However, after 25 years, many challenges have been raised to that original elegant and simple view (Merigan and Maunsell, 1993) (Hegde and Felleman, 2007), and an alternative description of the two pathways exists in terms of vision for perception (ventral stream) and vision for action (dorsal stream; Goodale and Milner, 1992). While the original model and its variant still serve as useful paradigms for interpreting results from psychophysics, neurophysiology, neuroanatomy, neuropsychology, and functional imaging, they are still evolving to incorporate newer findings. The objective of this article is to highlight a number of studies that together suggest the two pathways are functionally integrated in normal object recognition to enhance cue-invariant and viewpoint-invariant recognition by use of 3-D information. This may at first appear to contradict the original ideas of Ungerleider and Mishkin (1982) or those of Goodale and Milner (1992), but at closer inspection it will be evident that normal object recognition and all the variable viewing conditions that may challenge it necessitate the integrative action of these two streams.

First, the discussion will focus on the nature of object recognition in the ventral stream. It will be suggested that ventral object representations are largely viewpoint-invariant, although this invariance may not be represented at the single-cell level. Additionally, it will be suggested that familiarity with objects drives the development of representations that are more viewpoint-invariant. Finally, given that expertise with an object class requires extensive knowledge and familiarity with many members of the class, it is suggested that categories of objects that one has developed expertise in have a greater facility at achieving viewpoint-invariant representations for individual members of that class.

Second, the case is presented for dorsal–ventral integration in object recognition. Considering the primary discussion of viewpoint-invariant representations in the ventral pathway, it is suggested that shapes defined by 3-D cues that are dorsally extracted (particularly structure-from-motion) are ultimately processed by ventral stream mechanisms for recognition.

Through these two syntheses, it is proposed that normal object recognition likely requires the integrative action of the dorsal and ventral streams. This leads to several conjectures as to the properties of ventral stream representations, such as their invariance with respect to 3-D depth cues.