Towards a Robost Framework for Deep Person Reidentification

This paper aims to further improve the two key metrics of efficiency and accuracy of the existing pedestrian reidentification (ReID) method, based on which a more effective ReID model is proposed. Adding cutout data augmentation method to the original network backbone, embedding BN layers between network layers, and applying GEM (Generalized Mean Pooling) pooling method, experimental data show that the model has significantly improved both experimental results on Market1501 dataset, with 11.9% improvement in mAP and Rank-1 metrics respectively and 6.7%, respectively. The improvement in prediction efficiency and accuracy is achieved.