1 Introduction

The three most commonly used features in computer vision are regions, edges, and corners. Whereas regions are normally attributed properties to make them useful as input for subsequent processing stages such as matching, until recently edges and corners were described by little except for strength. However, over the past few years more attention has been paid to the properties of edges and corners too, since richer descriptions make them more effective features. For instance, in tracking, model matching, or model indexing, corner properties are capable of constraining the corner correspondences either as unary constraints or n-ary constraints between several corners. The latter enables viewpoint invariance. For example, under orthographic projection selected pairs of coplanar corners may be expected to have identical, but unknown, subtended angles or orientations. In [ 14 ] we demonstrated how the addition of relative colour and subtended angle corner properties (binary and unary constrains respectively) enabled the number of arcs in the association graph to be drastically reduced in a model recognition application using maximal clique graph matching.

Recent examples of edges properties being measured are scale [ 1 ], diffuseness, height and width [ 18 ]. Regarding corners, several interesting methods have been developed recently to measure scale, orientation, subtended angle, contrast, and blurring [ 2 , 5 , 7 , 8 , 11 , 12 ]. A problem with these techniques is that they tend to be complex and iterative, which impacts on both their efficiency and reliability. In contrast, rather that attempt the complicated task of simultaneously detecting and describing corners, in a previous paper we decoupled these two stages [ 14 ]. Standard algorithms were used to identify corner locations, and simple non-iterative methods were then sufficient for extracting corner properties. Many of the techniques we developed were based on local intensity or orientation histograms generated in the corner's neighbourhood. Before corner properties could be measured the histograms were smoothed by an automatically determined amount and the two main peaks located. A weakness of this approach is that it depends on correctly locating the peaks, which in turn depends on the appropriate level of smoothing being correctly determined. In this paper we develop further techniques for measuring corner properties more directly without recourse to histograms. By eliminating one potential source of error this has the potential for improving reliability and accuracy. In addition to the previously extracted corner properties we describe methods for measuring bluntness (or degree of rounding of the apex) and boundary shape (i.e. straight or curved sides).