Shape Representation

Abstract

After detecting the shape of an object, contour, or connected component, we should represent it in a concise and informative way. The simplest shape representation is a list of pixel coordinates. It can easy be obtained from the detection algorithm and includes precise spatial information. However, it is not ideal for further processing by higher-level algorithms for purposes such as object recognition or scene understanding. It does not characterize the shape in a useful way or indicate any shape features such as curvatures, slopes, or angles. Furthermore, this type of representation only describes a shape at a specific position, orientation, and scale. In summary, a list of coordinates describes a shape in a way that is too low-level and local to be immediately useful for high-level processing. We know that the early stages of the human visual processing stream use simple, local features for representing visual input. As information is passed on to higher levels, representations become more and more high-level, abstract, and location invariant. Ideally, our computer vision systems should do the same. It is would be highly inefficient to stick with pixel-based representations throughout the processing hierarchy and then attempt to accomplish the highest-level task based on such input. The highest level has to solve the most intricate problems and needs the lower levels to prepare the visual information in such a way that its tasks become computationally feasible. In this chapter, we will explore a variety of methods for representing shape at higher levels. Each of these techniques operates at a different level and emphasizes particular shape features. When choosing a representation for a computer vision system, we should of course always keep in mind the purpose of the system. Before starting any implementation, we should have developed a plan for all processing stages, their functionality, and the interfaces between them.