Evaluation of performance of grape berry detectors on real-life images

Abstract

Grape berry detectors based on SVM and HOG features have proven to be very efficient in detection of white grapes varieties. This statement is based on results, which have been achieved by 10-fold cross-validation, and by evaluation of the detectors on datasets with symmetrical prior probabilities of classes. The detectors have been also tested on real-life images; however, their performance could not be fully assessed in this case. The poor evaluation was caused by sensitivity of some of the used performance measures on composition of datasets. In order to obtain more useful results, all the used biased measures have been modified. The idea behind the modification, as well as the modification itself, is described in this paper. The modified measures have been used by re-evaluation of the detector's performance on a set of real-life images. The set had in fifteen real-life images, which were used within the original tests; however, this set has been extended to about thirty new images. The extended set allows obtaining of more precise information about performance of the detectors on real-life images. The results, which have been achieved by the re-evaluation, confirm expected excellent performance of the detectors on real-life images.