Recognition of curved shapes using geometric invariants

Planar shapes, which are images of 3D objects, are often used to identify the objects. In projective geometry, there are well-known techniques that are invariant to perspectivity and can be used to classify images obtained under different view angles. To apply these techniques corresponding points between two images (the model and the test image) have to be found. For objects with regular shapes such as quadrilaterals, corresponding points can easily be located. However, with curved shapes methods must be devised to locate a few of these points before applying further processes. We present a method to find prominent points of a curved shape by constructing a convex hull on the shape. Taking these points as references, cross ratios are computed as indices for comparing the test images with the models in database.