Iterative embedding distillation for open world vehicle recognition

Abstract

Vehicle recognition poses a practical but challenging problem in many real-world surveillance applications. Since vehicle recognition is an open-set problem, it is a critical issue to learn a discriminative visual embedding space rather than a well-performing classifier. In this paper, we propose an iterative embedding distillation (IED) framework for open-set vehicle recognition. The soft target in knowledge distillation is utilized to establish the interclass relations from an instance level rather than a category level. Towards the open-set problem, we extend knowledge distillation to embedding distillation in an iterative learning way, in which three types of loss functions are studied to iteratively transfer the distributions of embeddings from the teacher network to the student network. To demonstrate the universal nature of IED, we implement the IED framework on two basic convolutional neural networks and verify it using the cross-dataset testing protocols without retraining or fine-tuning. Extensive experimental results show that IED obtains quite encouraging results and outperforms state-of-the-art methods on various large-scale vehicle recognition datasets including VeRi-776, Vehicle-ID, Vehicle-1M, VD1 and VD2.