MEIAH: Mixing explicit and implicit formulation of attributes in binary representation for person re-identification

摘要

Person Re-identification (ReID) is an important yet challenging task in computer vision. It is far from solved due to the diverse background clutters, variations on viewpoints and body poses. On top of it, effective fast re-identification with binary representation is far more challenging. In this context, how to extract discriminative and robust binary features for identifying people in a large gallery is the core problem. It is observed that the pedestrian attribute labels can be good auxiliary information for learning better features for ReID task, but in most of the application scenarios we do not have the labeled training set with both pedestrian ID and attributes. In this paper, we first introduce a multi-task training method with data from target domain and auxiliary domain with different label types that is able to Mix Explicit and Implicit Attributes for Hashing (MEIAH). MEIAH is a novel end-to-end multi-task model to learn a mixed binary representation with explicit and implicit formulation of attributes for better ReID performance. Our architecture effectively unifies and takes full advantage of information from different domains. We evaluate the proposed method in four different bit lengths on two public benchmark datasets, including CUHK03 and Market-1501. Extensive experimental results show that the proposed method is effective and achieves the state-of-the-art results.