A Deep learning based evaluation method for feature detectors using satellite images

Feature detectors play an important role on computer vision. So far, feature-based and region-based methods, such as the repeatability, have been used to evaluate the performance of these detectors. Although these methods can provide an evaluation measure about the detection repeatability, they do not consider the informative part of the detected features as a whole set. In this work, we propose a novel method to evaluate feature detectors using a classification-based application. Kernel density estimation is used to represent the features sets and an effective SVM-based method is used to increase the number of variables on the feature space. An SVM classifier is then trained on the deep learning features which are generated using sparse autoencoders. Experimental results on satellite images show great difference among some common detectors and reveal the effect of their spaces on the classification accuracy.