Person retrieval in surveillance videos using attribute recognition

In person attribute recognition (PAR), an individual is described by his or her appearance. PAR-based person retrieval is a cross-modal problem where the input is a textual description of the person’s appearance and the output is an image of the person. The paper introduces PAR model development by merging a large-scale RAP dataset with the person retrieval benchmark dataset of AVSS 2018 challenge II. It uses a single deep network to detect a person’s attributes. The proposed approach uses five attributes; age, upper body (uBody) clothing color, uBody clothing type, lower body (lBody) clothing color, and lBody clothing type. Mask R-CNN is used for person detection, and the approach weighs each attribute to generate a ranking score for every detected person. Unlike the existing approaches, the proposed method uses a single deep network and fewer attributes to achieve state-of-the-art average intersection-of-union (IoU) of 66.7