A sample‐proxy dual triplet loss function for object re‐identification

Object re-identification, such as vehicle re-identification or pedestrian plays a significant role in intelligent video surveillance systems for public security. Due to viewpoint variations and appearance changes, both pedestrians vehicles usually have complex intra-class variations. However, most existing object methods often use sample-level triplet loss function cooperating with single-proxy softmax function, which could not handle well. In this paper, sample-proxy dual (SPDT) is proposed, works multi-proxy (MPS) function. The MPS charge of learning multiple proxies represent class. SPDT responsible enlarging inter-class distances well shrinking on sample proxy levels. Therefore, the method only handles but also fully learns discrimination samples proxies. Experiments two large datasets, that is, VeRi776 DukeMTMC-reID, demonstrate superior state-of-the-art approaches.