Constructing a visual memory in terms of uniqueness

Abstract

We propose a new scheme in representing and memorizing an image for pattern matching, scene understanding, and image database retrieval. A pattern in a visual space is evaluated in terms of uniqueness that reflects how stable and distinguishable the pattern is in its matching and recognition process. Features are accessed in the images in two fashions that are similar to saccade and smooth search in human vision, which makes the matching process to any particular feature be controllable. A view is organized in a structure linking unique patterns to less unique ones. Our uniqueness of a pattern is measured from both attributes of the pattern and its spatial relationship with others in its visual context. Establishing correspondence between two close views using this representation is not only efficient in accessing image primitives, but also reliable compared with conventional matching approaches that are based on restricted geometry constraints. It will be effective in active vision, memory based recognition, qualitative description and reasoning of scenes.