Perceptual Grouping and Shape Abstraction

点击翻译按钮获取中文摘要

Object categorization continues to be one of the most important challenges facing computer vision. When a particular target object is searched for in an image (the object detection problem), the target provides a strong shape prior to help segment and interpret the image features. But for the more general task of object categorization from a large database, no such target object is available. Instead, we must rely on a set of object-independent, mid-level shape priors that reflect the regularities of our world -- regularities identified by the Gestalt psychologists 100 years ago. The problem can be formulated as follows. Given an image of a cluttered scene and no a priori knowledge of scene content, our first task is to group together image features that belong to the same object. This is the classical problem of perceptual grouping, which has been largely ignored by the vision community in recent years due to the prominence of the object detection problem, in which stronger object-level priors subsume weaker mid-level priors. One of the primary objectives of this proposal is therefore to model these mid-level shape priors, such as symmetry, continuity, junctions, and closure, and to develop algorithms that can detect these regularities and use them to segment a scene into causally-related collections of image features, each corresponding to a different object. ******Ultimately, we would like to recognize the objects that these feature collections depict. But herein lies the problem when dealing with categories that exhibit high within-class variation. The granularity of the features comprising a categorical model is very coarse, whereas the granularity of the image features comprising a collection is very fine. This disparity is often referred to as the semantic gap in computer vision. We must therefore abstract (or regularize) the groups of local image features in order to "lift" them up to the level of the coarse, prototypical features that make up the models in the database. This raises many critical research questions also addressed in this proposal: 1) how do we model the abstract shape of an object?; 2) what are the invariant parts whose detection can help us determine what object we may be looking at? (a problem known as object indexing); 3) how can we use knowledge of a vocabulary of abstract parts to drive the abstraction process that will help us recover such parts from an image?; and 4) should these parts be 2-D or 3-D? ******Our research program addresses the two important and closely related problems of perceptual grouping and shape abstraction on multiple fronts. Without any knowledge of scene content, the ability to recover from an image a set of abstract, part-based image features yields a powerful indexing mechanism that can prune a large database down to a small number of promising candidates which, in turn, can provide strong, top-down priors to segment and detect the objects in a scene.**