Dynamic Pseudo Labels Sampling for Semi-Supersvied Person Re-Identification

Abstract

Person re-identification approaches based on deep learning have recently advanced considerably. However, most of the existing re-ID studies focus primarily on closed-set applications and neglect realistic, open-world settings; this limitation restricts the application of person re-ID in practice. Generally, closed-set re-ID methods usually assume that annotated data is sufficient for use when training or testing a deep learning model. However, in open-set applications, it is unrealistic to collect and annotate a available number of training data. For this purpose, we proposed a semi-supervised learning person re-identification method by dynamic pseudo-labelled data sampling. Using the dynamic pseudo labels sampling framework to fully exploit labelled and unlabelled data, we train and optimize the model in an iterative process. The semi-supervised learning method labels only a subset of pedestrians, while the single-sample learning method labels only one data point for each pedestrian. Then, a pseudo label is generated for the unlabelled data, and the labelled data and a subset of the pseudo-labelled data are then used as an extended dataset for training. To make use of the unlabelled data, pseudo labels are generated from the unlabelled data with a dynamic strategy. Based on the predicted confidence, only use selected pseudo-labelled data for training, and the Convolution Neural Network (CNN) model is updated through retraining with an extended dataset. The effectiveness of this method is demonstrated by experiment on publicly available datasets. Compared with the state-of-the-art methods, the performance is improved.