Reliable shape-based recognition of objects in images

 People understand an unexpected new image amazingly fast; how can an
artificial recognition system do the same? As a first step, it must
find image attributes distinctive enough to reduce the image's
possible interpretations to a relatively small, manageable number, and
researchers agree that this requires perceptual organization (PO), the
identification of mid-size image structures such as salient lines,
regions or object shapes. Yet the intrinsic ambiguities of images make
PO unreliable, and recognition systems based on PO have not worked
well.

 I present a recognition method that exploits PO without succumbing to
 its unreliability. The method recognizes objects by shape. It
 compares the shapes in two images by comparing the two images'
 segmentations (i.e., their partitions into homogeneous regions) as
 computed by PO, using a new similarity measure based on mutual
 information. To overcome the segmentations' unreliability, we compute
 the average similarity between all possible segmentations of the two
 images weighted by probability, in a simple, accurate closed-form
 approximation. The result is a robust measure which directly compares
 images according to their contained shapes, without needing to find
 the shapes explicitly.  Other applications include tracking moving
 objects in video, computing stable image segmentations, and
 edge-preserving smoothing.