This paper evaluates the performance of our corner detector and four popular corner detectors under similarity and affine transforms. The majority of authors of published corner detectors have not used theoretical criteria to measure the stability and accuracy of their algorithms. They usually only illustrate their results on different test images and compare them to the results of other test corner detectors. A few of them use only one criterion. This criterion is the number of matched corners between original and transformed images, divided by the number of corners in the original image. This criterion is flawed since it favours algorithms which find more false corners in input images. We propose two new criteria to evaluate the performance of corner detectors. Our proposed criteria are "consistency of corner numbers" and "accuracy". These criteria were measured using many images and experiments such as rotation, uniform scaling, non-uniform scaling and affine transforms. To measure accuracy, we created ground truth based on majority human judgement. The results show that our corner detector performs better under similarity and affine transforms. Keywords:corner detection, curvature scale space, consistency, accuracy, similarity and affine transforms.
Home Contents Author index Keyword index