Pseudo-label estimation via unsupervised Identity Link Prediction for one-shot person Re-Identification

In this paper, we propose an unsupervised identity link prediction (ILP) method for label estimation in one-shot person Re-ID. ILP aims to relax the constraints of labeled samples and group a set of unlabeled pedestrians by their potential identities. The main idea is that the category relationships between pedestrians (nodes) can be inferred from their local context in the feature space. Specifically, an Identity Link Subgraph (ILS) describes the link relationship between nodes and their nearest neighbors, which is constructed by a two-step procedure. Meanwhile, a Dynamic Penalty Module (DPM) is introduced at each ILS construction step to infer which linkage between pairs in the ILS should be pruned to assign higher-quality classification labels. To fully use the accurate identity information in initial labeled samples, we jointly use identity pseudo-labels (which are estimated by adopting the Nearest Neighbors classifier) with classification pseudo-labels for model training. Moreover, we design a Dual-Branch Fusion network (DBF-Net) to optimize the CNN model simultaneously through all (pseudo-)labeled samples. Results on multiple datasets prove that DBF-Net outperforms traditional one-shot Re-ID methods by a large margin.