Curvature can be used in computer vision for e.g. object recognition or motion and pose estimation. Methods have been developed to estimate curvature in 2D images but few have been devloped for arbitrary 3D surfaces.
In Geometry an overview of the geometry involved is presented, in Curvature Estimation Techniques three methods for choosing the curvature extrema (corners) are considered as examples of the application of these ideas, and finally a list of references are presented.
Given a point on a 3D surface there are two principal descriptions of the local curvature, the curvature class (the possible types are given in Table 1) and the curvature magnitude. The curvature class and the curvature magnitude can be computed from the two principal curvatures κ1 (the maximum local curvature) and κ2(the minimum local curvature).
The curvature of a surface at a given point is a measure of the deviation from the tangent (see Equation 1); the normal curvature is the curvature in the plane of the normal.
Figure 1(a) shows a cylinder intersected by a normal plane (N) in n (the normal vector to the point P) and u (the tangetial vector to the point P). u must be orthoganol to n but can can take on any direction in that plane (0° to 360°). The normal curvature is plotted in Figure 1(b) and from this κ1 and κ2 can be computed. In this example the minimum curvature (C2) is a straight line with κ = 0 and the maxminum curvature (C2) κ = 1 * radius of the circle (see Figure 1(c)).
In Figure 2 the two normals n1 (=n) and n2 (=u) are seperated by a distance L on the surface of the object. The curvature κcan be computed from n1, n2 and L. L must be computed; this is different for different geometries (e.g. Figure 3 (a) shows a cylinder, Figure 3 (b) an Elipsoid).
The mean curvature H (see Equation 3) and the Gaussian curvature K (see Equation 2) can be calculated from κ1 (p1) and κ2 (p2).
The surface is modelled as a series of adjacent triangles. Each triangle is flat so the curvature is concentrated at the common vertex of the triangles. Equation 4 defines the curvature when using this method.
A cross shaped pair of points that are in two orthogonal planes are sampled and assumed to be of polynomial form. The coefficents of the polynomial can be used to estimate the surface curvature.
A series of smoothing filters are applied to the image; the surface curvature can be computed from first and second order partial derivatives of the smoothed image.
Images are referenced in mouse-overs from the Further Information section.