| We present a new computational framework for solving the perceptual organization problem in vision.; Humans have a remarkable ability to organize their visual input. Instead of seeing a collection of values associated with individual photoreceptors, we perceive a number of visual groups, usually associated with objects or well-defined parts of objects. This ability is equally important for computer vision. To recognize objects, we must first separate them from their backgrounds; manipulation and navigation requires knowing what and where the objects are in a scene.; Grouping is a complex problem with many intertwined issues and subgoals. Several key aspects of grouping are: (1) multiple cues are used--the Gestaltists have identified many grouping factors, such as proximity, similarity, good continuation, common fate, symmetry, most of which can be considered to be similarity in some feature space, (2) grouping is done through a global decision process--the Gestaltist uses the notion of "Pragnanz" to express the aim of grouping as extracting "holistic impression" of an image through global interactions within the visual system, (3) multiple level of interpretation and grouping--grouping is not just about clustering in a low level feature space, it is also about tapping into prior knowledge of the world to bring a more meaningful description to the image.; Here we propose a new framework for grouping which aims to capture these aspects of grouping. Grouping is viewed as a process of (1) defining the similarity between the image primitives, and (2) a generic grouping engine which groups the image primitive together based on the feature similarity measure. The proposed grouping engine, Normalized Cuts, formulates grouping as a hierarchical graph partition problem. The input to the algorithm is the similarity measure between the image primitives. Based on the similarity measure, a weighted graph is constructed. The grouping criterion, the normalized cut, measures both the total dissimilarity between the different groups as well as the total similarity within the groups. An efficient algorithm is developed for finding optimal partition by solving a generalized eigenvector system. We applied this algorithm to segmenting real complex images based on brightness/color, texture, and motion, and found results very encouraging. |
