Person Re-Identification Using Hybrid Representation Reinforced by Metric Learning

Abstract

Person Re-Identification (Re-Id) is among the main constituents of an automated visual surveillance system. It aims at finding out true matches of a given query person from a large repository of non-overlapping camera images/videos. In this paper, we have proposed an efficient Re-Id approach that is based on a highly discriminative hybrid person representation which combines the low-level hand-crafted appearance based features together with the mid-level attributes and semantic based deep features. The lowlevel hand crafted features are extracted by using hierarchical Gaussian and local histogram distributions in different color spaces. These features incorporate discriminative texture, shape and color information which is invariant to distractors, e.g., variations in pose, viewpoint and illumination, and so on. The mid-level attribute based deep features are extracted to incorporate contextual- and semantic-based information. The feature space is optimized and self-learned using cross-view quadratic discriminant analysis and multiple metric learning, with the aim to reduce the intra-class differences and increase the inter-class variations for robust person matching. The proposed framework is evaluated on publicly available small scale (VIPeR, PRID450s, and GRID) and large scale (CUHK01, Market1501, and DukeMTMC-ReID) person Re-Id datasets. The experimental results show that the hybrid hand-crafted and deep features outperformed the existing state-of-the-art in approaches in the unsupervised paradigm.