Global-to-Local Dynamic Feature Aggregation for Unsupervised Person Re-Identification

Most of existing unsupervised person re-identification algorithms use single-scale structure to extract global features. However, the features extracted from different scales and spatial locations are crucial for person re-identification task mainly using body parts to identify. In this paper, we propose a novel global-local architecture called dynamic feature aggregation network (DFANet) to learn discriminative patch features at various semantic levels on unlabelled datasets and aggregate them with a dynamic aggregation mechanism. Specifically, DFANet starts with a global feature learning stage to learn global features at various semantic levels. Then a local stage formed by attention patch generation network and dynamic aggregation module is deployed to extract distinct and notable local features on unlabelled datasets. Extensive experimental results on Market1501 and DukeMTMC show that our proposed method outperforms state-of-the-art works.