The human visual system can spot an abnormal image, and reason about what makes it strange. This task has not received enough attention in computer vision. In this paper we study various types of atypicalities in images in a more comprehensive way than has been done before. We propose a new dataset of abnormal images showing a wide range of atypicalities. We design human subject experiments to discover a coarse taxonomy of the reasons for abnormality. Our experiments reveal three major categories of abnormality: object-centric, scene-centric, and contextual. Based on this taxonomy, we propose a comprehensive computational model that can predict all different types of abnormality in images and outperform prior arts in abnormality recognition.

We present a probabilistic approach to shape decomposition that creates a skeleton-based shape representation of a 3D object while simultaneously decomposing it into constituent parts. Our approach probabilistically combines two prominent threads from the shape literature: skeleton-based (medial axis) representations of shape, and part-based representations of shape, in which shapes are combinations of primitive parts. Our approach recasts skeleton-based shape representation as a mixture estimation problem, allowing us to apply probabilistic estimation techniques to the problem of 3D shape decomposition, extending earlier work on the 2D case. The estimated 3D shape decompositions approximate human shape decomposition judgments. We present a tractable implementation of the framework, which begins by over-segmenting objects at concavities, and then probabilistically merges them to create a distribution over possible decompositions. This results in a hierarchy of decompositions at different structural scales, again closely matching known properties of human shape representation. The probabilistic estimation procedures that arise naturally in the model allow effective prediction of missing parts. We present results on shapes from a standard database illustrating the effectiveness of the approach.

Figure/ground assignment, in which the visual image is divided into nearer (figural) and farther (ground) surfaces, is an essential step in visual processing, but its underlying computational mechanisms are poorly understood. Figural assignment (often referred to as border ownership) can vary along a contour, suggesting a spatially distributed process whereby local and global cues are combined to yield local estimates of border ownership. In this paper we model figure/ground estimation in a Bayesian belief network, attempting to capture the propagation of border ownership across the image as local cues (contour curvature and T-junctions) interact with more global cues to yield a figure/ground assignment. Our network includes as a nonlocal factor skeletal (medial axis) structure, under the hypothesis that medial structure ``draws'' border ownership so that borders are owned by their interiors. We also briefly present a psychophysical experiment in which we measured local border ownership along a contour at various distances from an inducing cue (a T-junction). Both the human subjects and the network show similar patterns of performance, converging rapidly to a similar pattern of spatial variation in border ownership along contours.

We present an account of human concept learning-that is, learning of categories from examples-based on the principle of minimum descrip(cid:173) tion length (MDL). In support of this theory, we tested a wide range of two-dimensional concept types, including both regular (simple) and highly irregular (complex) structures, and found the MDL theory to give a good account of subjects' performance. This suggests that the intrin(cid:173) sic complexity of a concept (that is, its description -length) systematically influences its leamability. 1- The Structure of Categories A number of different principles have been advanced to explain the manner in which hu(cid:173) mans learn to categorize objects. It has been variously suggested that the underlying prin(cid:173) ciple might be the similarity structure of objects [1], the manipulability of decision bound~ aries [2], or Bayesian inference [3][4]. While many of these theories are mathematically well-grounded and have been successful in explaining a range of experimental findings, they have commonly only been tested on a narrow collection of concept types similar to the simple unimodal categories of Figure 1(a-e).

This paper summarizes a novel logic-based approach to grouping and perceptual organization, and presents novel efficient methods for computing interpretations in this framework. Grouping interpretations are first defined as logical structures, built out of atomic premises ("regularities") that are derived from considerations of non-accidentalness. These interpretations can then be partially ordered by their degree of regularity or constraint (measured numerically by their codimension). The genericity constraint-the principle that interpretations should minimize coincidences in the observed configuration-dictates that the preferred interpretation will be the minimum in this partial order, i.e. the interpretation with maximum codimension. The preferred interpretation, called the qualitative parse, corresponds neatly to the interpretation intuitively preferred by human observers. As a side-effect, the "most salient" or most structured part of the scene can be identified, as the highest-codimension subtree of the qualitative parse. An efficient (O(n/sup 2/)) method for computing the maximum codimension interpretation is presented, along with examples.

This paper investigates perceptual grouping from a logical point of view, defining a grouping interpretation as a particular kind of logical expression, and then developing an explicit inference theory in terms of such expressions. First, a regularity‐based interpretation language is presented, in which an observed configuration is characterized in terms of the regularities(special configurational classes, e.g. nonaccidental properties) it obeys. The most preferred interpretation in such a system is shown to be the most‐regular (maximum “codimension”) model the observed configuration obeys, which is also the unique model in which it is generic(typical). Inference then reduces to a straiightforward exercise in LogicProgramming. Because generic model assignment involves negation, this reduction requires that a version of the Closed World Assumption (CWA) be adopted. Next, this entire regularity‐based machinery is generalized to the grouping problem: here an interpretation is a hierarchical recursive) version of a model called a parse tree.For a given number of dots and a fixed choice of regularity set, it is possible to explicitly enumerate the complete set of possible grouping interpretations, partially ordered by their degree of regularity (codimension). The most preferred interpretation is the one with maximum codimension (i.e., the most regular interpretation), which we call the qualitative parse. An efficient procedure (worst case $O(n^2)$) for finding the qualitative parse is presented. The qualitative parse has a unique epistemic status: given a choice of regularity set, it is the only grouping interpretation that both (a) is maximally regular, and (b) satisfies the CWA. This unique status, it is argued, accounts for the perceptually compelling quality of the qualitative parse.

It is shown how a parameterization of the structure of the observed object (interpretable as the space of dimensions of generative operations which brought the object into existence) entails a lattice that enumerates all the allowable categories and subcategories for the class of object, along with an inferential preference hierarchy among them. Each model is constrained to be generic in its parameterization, so that each node in the lattice stands in for an entire class of objects that, all being parameterized the same way, can all be treated as equivalent to one another: the observed object's natural category.>