Center based pseudo-labeling for semi-supervised person re-identification

Generative Adversarial Networks (GAN) have shown promising results on data modeling and can generate high quality synthetic samples from the data distribution. However, how to effectively use the generated data for improved feature learning still remains an open question. This work proposes a Center based Pseudo-Labeling (CPL) method dedicated to this purpose. The network is trained with both labeled real data and unlabeled synthetic data, under a joint supervision of cross-entropy loss together with a center regularization term, which simultaneously predicts pseudo-labels for unlabeled synthetic data. Experimental results on two standard benchmarks show our approach achieves superior performance over closely related competitors and comparable results with state-of-the-art methods.