Disentangled Feature Learning for Semi-supervised Person Re-identification

Jin Ding, Xue Zhou, Gehan Hao

Abstract: Semi-supervised learning has become more and more popular in person re-identification because acquiring annotations is time-consuming and cumbersome. In this paper, we design a Disentangled Feature Learning (DFL) model based on an encoder-decoder mode by making use of less labeled data together with lots of unlabeled data. Specifically, a pair of horizontally flipped images are fed into the encoder module to decompose identity and structure features. Then different decomposed features are combined to reconstruct images in the decoder module. Multi-view consistent constraints are proposed for the further learning of feature disentanglement in the image-view, feature-view and feature-vector-view, respectively. Moreover, we separate the identity encoders into a teacher encoder and a student encoder for the stability of training and better performance. Extensive experiments on four ReID datasets demonstrate the effectiveness of our DFL model in the low proportion of the labeled data.