Evaluating the angular sensitivity of corner detectors

Several popular corner detectors were evaluated on imagery containing corners with a variety of internal angles. Even in a noise-free environment, differences in performance were found. A null hypothesis approach was taken in evaluating whether these performance differences were significant, taking into account correctly the size of the dataset and the number of discrepancies. It was found that some of these performance differences are statistically significant, allowing recommendations to be made regarding which detectors should be used when a problem has corners of known internal angles.