No abstract

Person re-identification technology is an important foundation for security, pedestrian tracking and other fields, and is the key to building a safe and smart city. In recent years, a large number of researchers have trained pedestrian re-identification networks through triplet loss, especially the triplet loss with batch hard mining (BHTri loss) has greatly improved person re-identification networks on accuracy. However, triplet loss with batch hard mining for an anchor sample only select the hardest positive sample and the hardest negative sample to calculate the loss, ignoring the influence of other samples on network parameters. In response to the above issues, this paper proposes a variant of triplet loss with batch hard mining, which is called adaptive weight triplet loss with batch hard mining. After the training dataset extracts features from the backbone network, in the phase of calculating loss, it takes the average of the sum of the distances between an anchor sample and all corresponding positive samples as a threshold, and both positive samples with an anchor point distance greater than the threshold and negative samples smaller than the threshold are retained, then based on The distance of the anchor sample is given the corresponding sample weight for calculating the loss. Compared with BHTri, mAP have improved 1.79%, 2.04%, and 1.25% respectively on the Market1501, DukeMTMC-reID and CUHK03 datasets, indicating that the proposed algorithm is effective.