A novel performance evaluation method of local detectors on non-planar scenes

This paper presents a new method for performance evaluation of local detectors on non-planar scenes. The framework extends the detector evaluation of Mikolajczyk and Schmid to non-planar scenes. The basic idea for establishing ground truth information on 3D scenes is the use of the trifocal tensor and the point transfer property. The repeatability and matching scores (using the SIFT descriptor) for the currently most popular local detectors are compared for increasing viewpoint change. In addition, detector combinations are evaluated by estimating the diversity of regions produced by the different detectors. The experiments reveal that the results differ from that of previous evaluations on planar scenes, especially that repeatability and matching scores will be much lower in real applications as suggested by previous evaluations.