Global Based Deep Refineing Model For Person Retrieval

The performance of traditional part model in person retrieval is greatly affected by the quality of parts. The recent work[1] consider refining the hard partitioned part when training the network itself and got state-of-the-art performance, but during our experimentation, we found the masks which it generated contains the problems that misguide the networks with additional constraints, Targeting to solve above problem, we proposed a new networks called Global Refine Net. The backbone network focus on learning the local information which improve the ability of extract feature of details, Global Refine block introduce global information to adjust the hard-shaped part generated by the backbone network in an end-to-end manner. Also we modified the self-adversarial training mechanism in [1]. We employ an special loss function to prevent the incorrect convergence and adjust the degree of self-adversarial training, the new regularization term we added in the loss benefit both in stabilizing and speeding the training process. The performance of our model beat most previous soft partitioned works, improved about 2.3% rank-1 accuracy and 5.1% mAP to the PCB baseline on market-1501 dataset.