Multi-information Constraint Learning for Unsupervised Domain Adaptive Person Re-identification

Person re-identification (ReID) aims at identifying the same person’s images across different cameras. However large domain gaps between source and target domains, as well as lack of label information in the target domain poses a huge challenge for unsupervised domain adaptive the ReID model. This paper tackles the challenge through three aspects : (1) we design a robust visual-spatiotemporal fusion model, which improves the quality of pseudo labels based on visual probability evaluation, spatiotemporal probability evaluation and visual-spatiotemporal fusion. (2) we propose a novel sampling strategy for deep mutual information estimation and maximization algorithm (DIM), which employs data augmentation and dynamic storage stack to improve the reliability of the selected samples. (3) We combine the DIM loss and other supervised losses together to construct a new multi-objective loss function and present a corresponding dynamic adjustment strategy for the weights of loss functions, which contribute to stable the convergence of the training process. As shown in experimental results, our model has achieved excellent results on two ReID datasets, Market-1501 and DukeMTMC-ReID.