Viewpoint invariant subject retrieval via soft clothing biometrics

Abstract

As much information as possible should be used when identifying subjects in surveillance video due to the poor quality and resolution. So far, little attention has been paid to exploiting clothing as it has been considered unlikely to be a potential cue to identity. Clothing analysis could not only potentially improve recognition, but could also aid in subject re-identification. Further, we show here how clothing can aid recognition when there is a large change in viewpoint. Our study offers some important insights into the capability of clothing information in more realistic scenarios. We show how recognition can benefit from clothing analysis when the viewpoint changes with partial occlusion, unlike other approaches addressing soft biometrics from single viewpoint data images. This research presents how soft clothing biometrics can be used to achieve viewpoint invariant subject retrieval, given a verbal query description of the subject observed from a different viewpoint. We investigate the influence of the most correlated clothing traits when extracted from multiple viewpoints, and how they can lead to increased performance.