A statistical approach for comparing the performances of corner detectors

Corner detectors are widely used in computer vision. This paper assesses several state-of-the-art corner detectors in terms of overall performance and the internal angles of corners using simple geometric shapes. This assessment is carried out using a statistically-valid null hypothesis approach, not previously used in computer vision. It is found that there are statistically significant differences in performance. Moreover, the null hypothesis approach is easy to use in comparing vision techniques.