APPROACH TO LANDMARK-BASED SHAPE RECOGNITION.

We introduce a new method of comparing two-dimensional shapes based on landmarks or tokens derived from the shape contour. This approach is motivated by the difficulties associated with shape recognition of an object when the segmentation process fails to produce the entire object contour (or silhouette) due to errors in the segmentation algorithm or object occlusion. We derive a measure of similarity for objects with three landmarks. For objects with more than three landmarks, we develop a method based on two statistical shape measures and another method using a MMSE (minimum mean square error) affine transformation to differentiate shapes.