Person Re-Identification based on Data Prior Distribution

In order to solve the problem of insufficient accuracy of the existing person re-identification methods. We propose a neural network model for identifying pedestrian properties and pedestrian ID. Compared with the existing methods, the model mainly has the following three advantages. First, our network adds extra full connection layer, ensure model migration ability. Second, based on the number of samples in each attribute, the loss function of each attribute has been normalized, avoid number unbalanced among the attributes to effect the identification accuracy. Third, we use the distribution of the attribute data in the prior knowledge, through the number to adjust the weight of each attribute in the loss layer, avoid the number of data sets for each attribute of positive and negative samples uneven impact on recognition. Experimental results show that the algorithm proposed in this paper has high recognition rate, and the rank-1 accuracy rate on DukeMTMC dataset is 72.83%, especially on Market1501 dataset. The rank-1 accuracy rate is up to 86.90%.