An Efficient Person Reid Methed Based on Knowledge Distillation

Existing Person Reid (Person Re-identification) approaches often only consider how to improve the model generalization performance, which unfortunately comes at the cost of massive computation and storage consumption. Despite that those extensive neural networks are powerful to deal with a specific task and achieve excellent results on many tasks, they are too large to be deployed on edge devices like smart-phones and security cameras. Recently, [5] (Hinton, Vinyals, and Dean 2014) have shown that the dark knowledge within a powerful and massive neural network (a.k.a. teacher) can significantly help the training of a smaller and faster student network (a.k.a. student) and improve performance of the student network. In this work, we investigate the knowledge distillation strategy for training small Person Reid networks by making use of large networks. To this end, we present the Efficient Person Reid model, which is achieved by effective knowledge transferring the dark knowledge of a powerful teacher network. Extensive experiments on Person Reid standard datasets DukeMTMC substantiates the effectiveness of our proposed approach. Particular our efficient Reid model achieves Rank-1/mAP = 86.8%/72.9% on DukeMTMC with only 256 dimensions' feature vector and with only 3.4M Params.