Localization Accuracy of Region Detectors

Abstract

In this paper, a comparison of five state of the art region detectors is presented with regard to localization accuracy in position and region shape. Based on carefully estimated ground truth homographies, correspondences between frames are assigned using geometrical region overlap. Significant differences between detectors exist, depending on the type of images. Also, it is shown that localization accuracy linearly depends on region scale for some detectors, which may thus be used as a pre-selection criterion for the removal of error-prone regions. The presented results serve as a supplement to existing comparative studies, and can be used to facilitate the selection of an appropriate detector for a specific target application. When descriptor distance is used as assignment criterion instead of region overlap, a different set of correspondences results with lower accuracy. Set differences (and thus localization accuracy) are directly related to the density of regions in a local neighborhood. Based on the latter, a novel measure for the identification of error-prone regions - shape uniqueness - is introduced. In contrast to existing methods that are based on the descriptor distance of region correspondences, the new measure is pre-computed on each image individually. Thus, the complexity of the subsequent matching task can be significantly reduced.