On evaluating methods for recovering image curve fragments

Yuliang Guo, Benjamin B. Kimia

Abstract: Curve fragments, as opposed to unorganized edge elements, are of interest and use in a large number of applications such as multiview reconstructions, tracking, motion-based segmentation, and object recognition. A large number of contour grouping algorithms have been developed, but progress in this area has been hampered by the fact that current evaluation methodologies are mainly edge-based, thus ignoring how edges are grouped into contour segments. We show that edge-based evaluation schemes work poorly for the comparison of curve fragment maps, motivating two novel developments: (i) the collection of new human ground truth data whose primary representation is contour fragments and where the goal of collection is not distinguished objects but curves evident in image data, and (ii) a methodology for comparing two sets of curve fragments which takes into account the instabilities inherent in the formation of curve fragments. The approach compares two curve fragment sets by exploring deformation of one onto another while traversing discontinuous transitions. The geodesic paths in this space represent the best matching between the two sets of contour fragment. This approach is used to compare the results of edge linkers on the new contour fragment human ground truth.