Teacher-Student Based Domain Adaptation for Person Re-identification

Abstract

In traditional deep Convolutional Neural Network (CNN) based person re-identification (Re-ID) methods, there must be thousands of training samples with annotated information under the same data distribution, in other words, the data used to train the CNN model is collected from the same scene. However, if we need to deploy a trained Re-ID model at the airport, but the data for training the model is collected on campus, the traditional methods will suffer considerable performance degradation because the data come from different distributions. It is a non-trivial task to straightforwardly re-collect sufficient annotations from the new scene, due to the high cost of time and human labor. Therefore, the domain adaptation paradigm is used to resolve the data distribution problem. Specifically, an efficacious method for domain adaptive person Re-ID based on the teacher-student framework is proposed in our research. Our method has been evaluated by numerous experiments and proved its effectiveness.