An improved baseline for person re-identification

Abstract

Person re-identification(Re-ID) using deep learning has made great progress in the past few years, but there is one problem that many state-of-the-art Re-ID methods all use a complex network most of which use the structure of multi-branch and multi-loss function. At present, the database used for Person re-identification is relatively small. This complex network structure may bring a problem that although current methods may perform well in the small databases, but there may be some problems of overfitting problem, once applied in the bigger dataset or real scene these complex methods may perform not well. So this paper mainly proposes a new powerful baseline network. This end-to-end network only uses a global feature and does not use multi-branch structure, but achieves state-of-the-art level. The key point is that this network has good improvement potential to adapt to larger datasets and even practical application scenarios.