Multi-scale Feature Fusion Network for Person Re-identication

In view of the similarities between the appearance of the object to be identified and the color of the object in the person re-identification, the problem with the model is that it is difficult to correctly identify the pedestrian's identity. To solve this problem, a multi-scale feature fusion network based on the residual network ResNet50 was proposed, in which the last layer feature was used to cooperate with multiple intermediate layer features. The top-level to lower-level progressive addition feature layer fusion mechanism was adopted, and the pattern of the pedestrian image ensured that the ability to represent small details in this mode was improved based on the overall feature representation. Experiments were carried out on three mainstream person re-identification public datasets Market-1501, CUHK03(D) and DukeMTMC-reID respectively. Compared with the same type of person re-identification network DaRe in 2018, the proposed method in Rank-1 indicator on the 1501 dataset increased by 2.82%, and the mAP indicator increased by 4.32%; the Rank-1 indicator of the DukeMTMC-reID dataset increased by 5.45%, and the mAP indicator increased by 6.4%. The experimental results show that the proposed method is effective.