View-based representations for dynamic 3D object recognition

Much of the experimental and computational modeling research on human recognition processes has focused exclusively on the domain of static three-dimensional (3D) objects. The issue of the nature of internal representations underlying dynamic 3D object recognition is largely unexplored. Here we examine this issue, with emphasis on view-point dependency, using variants of biological motion sequences of the kind described by Johansson (1973). Our first experiment investigated whether observers exhibit the well-known canonical view-point effect while recognizing 3D biological motion sequences. Results showing a markedly impaired recognition performance with sequences recorded from unusual view-points provide preliminary evidence for the role of view-point familiarity and the inability of the visual system to extract view-independent representations. Next, to examine whether the motion traces used for recognition preserve 3D information, or are largely 2D, we developed a special class of biological motion sequences. The distinguishing characteristic of these sequences was that while they preserve the `normal' 2D projections from one view-point, their 3D structures were randomized. View-points preserving the `normal' 2D projections yielded vivid biological motion percepts, whereas other viewpoints yielded percepts of randomly moving dots. In the final set of experiments we examined whether this result could be an outcome of a recognition-dependent top-down suppression of anomalies in 3D structures. Our results indicate that subjects' expectations about 3D structure can suppress the bottom-up depth information provided by binocular stereo. Taken together, these findings suggest that biological motion sequences are represented by the human visual system as 2D traces rather than as 3D structural descriptions, and that the perception of 3D structure may be based not only upon low-level processes but also upon recognition-dependent top-down influences