We present a flexible curve matching algorithm which performs qualitative matching between curves that are only weakly similar. While for model based recognition it is sufficient to determine if two curves are identical or not, for image database organization a continuous similarity measure, which indicates the amount of similarity between the curves, is needed. We demonstrate how flexible matching can serve to define a suitable measure. Extensive experiments are described, using real images of 3D objects. Occluding contours are matched under partial occlusion and change of viewpoint, and even when the two objects are different (such as the two side views of a horse and a cow). Using the resulting similarity measure between images, automatic hierarchical clustering of an image database is also shown, which faithfully capture the real structure in the data.