Guidelines for Query and Gallery Image Extraction in Person Re-Identification Systems

Abstract

Variations in visual factors such as camera viewpoint and background, person orientation and pose are commonly regarded as significant challenges in Person Re-Identification (Re-Id). In this study, we analyse how individuals’ orientation in query images selected by the operator and in gallery images automatically extracted by pedestrian detectors/trackers affects the performance of Re-Id systems. Since existing benchmark data sets of real images are not suitable for our analysis, we used two synthetic data sets. Our experiments reveal that lateral orientations are more informative than frontal ones, contrary to intuitive expectations, whereas they confirm that posterior orientations are the least informative. Our findings underscore the significance of individuals’ orientation for the accuracy of Re-Id systems and provide insights for both designers and users of Re-Id systems into effective selection criteria of the gallery and query images.