COMPARING MULTIPLE CORNER DETECTION ALGORITH

Abstract

Performance characterization of algorithms has been commonly performed using parametric methods such as Precision-Recall, Repeatability, and detection rate etc. These methods assume that the data to be normally distributed, and therefore, the results became data specific. The main objective of this analytical study was to employ non-parametric statistical tests, for this purpose two tests Wilcoxon Signed Rank test and McNemar’s test were applied to characterize the performance of corner detection algorithms. The results showed that the use of sufficiently large amount of data and correct testing framework using different non-parametric statistical tests yielded similar results, which was not observed with conventional parametric tests. Both Wilcoxon Signed Rank test and McNemar’s test produced a similar ranking of corner detection algorithms, as both tests suggested Harris and Stephens at the top position then SUSAN, FAST, GLC and finally KLT. Hence, these non-parametric test were recommended to be used for the evaluation of vision algorithm due to their simplicity and reliability.