Perceptual organization as generic object recognition

We approach some aspects of perceptual organization as the process of fitting generic models of objects to image data. A generic model of shape encodes prior knowledge of what shapes are likely to come from real objects. For such a model to be useful, it must also lead to efficient computations. We show that models of shape based on local properties of objects can be effectively used by simple, neurally plausible networks, and that they can still encode many perceptually important properties. We also discuss the relationship between perceptual salience and viewpoint invariance. Many gestalt properties are the subset of viewpoint invariant properties that can be encoded using the smallest possible sets of features, making them the ecologically valid properties that can also be used with computational efficiency. These results suggest that implicit models of shape used in perceptual organization arise from a combination of ecological and computational constraints. Finally, we discuss experiments demonstrating the role of convexity in amodal completion. These experiments point out some of the limitations of simple local shape models, and indicate the potential role that the part structure of objects also plays in perceptual organization.