Application of Knowledge Distillation in Representation Learning Person Re-identification Model

Chang Liu ； Hang Ma ； Jun-Jie Jin ； Xin-Lun Zhou ； Wen-Bai Chen

《電腦學刊》 31卷2期 (2020/04) Pp. 277-286

https://doi.org/10.3966/199115992020043102023

摘要

The lock targets in monitoring system is important for fully exerting the surveillance capability of mobile devices and saving working time. To save the time required and huge amount of computing resources, a fast person re-identification (Re-ID) method is proposed. In this paper, we use knowledge distillation to make a large teacher model (ResNet50) guide a small but effective student model (MobileNet v2) for representation learning. Experimental results demonstrate that the proposed method is feasible. Compared with the teacher model and the student model, the system applied the knowledge distillation method can save more 55.4% of time and increase mAP 12.73% and Rank-1 8.63%, respectively.

關鍵字

deep neural network ； knowledge distillation ； person Re-ID ； representation learning