Abstract
The role of 3D information in object recognition is a fundamental unresolved issue in high-level vision. Here we explore whether recognition can operate on exclusively 3D cues and how 3D cues interact with 2D information. Superficially, the first issue seems to admit a ready answer — given that we can recognize objects in random-dot stereograms, where the cues are putatively purely 3D, one may conclude that 3D information is sufficient for recognition. However, this conclusion can be challenged on the grounds that recognition in this situation may be based not on the perceived 3D information per se, but rather on the 2D distribution of depth discontinuities and iso-depth contours. In our first experiment, we examined whether there were any differences in recognition performance (accuracy and reaction time) when objects in random-dot stereograms had correct 3D structure versus incorrect 3D shape that precisely preserved the 2D distribution of the depth discontinuities and iso-depth contours. We found a significant difference between the two conditions. On a set comprising ten common objects such as faces, vehicles and animals, subjects' recognition accuracy with incorrect 3D information was significantly reduced relative to recognition accuracy with the correct 3D structures. Next, we examined the interaction between 3D and 2D cues by testing whether the inclusion of 3D cues could shift the threshold luminance contrast needed for recognition. Our data so far do not reveal any such shift, suggesting that the analysis of 2D and 3D cues for recognition may be relatively independent. Planned future experiments include the use of more sensitive methods, such as progressive deblurring, for detecting shifts in the onsets of object recognition across the correct and incorrect 3D conditions, and imaging studies for determining whether common cortical areas subserve object perception based on 2D versus purely 3D cues.