FP-GCN: fine pseudo-label driven iterative GCN to learning discriminative fusion features for unsupervised person re-identification

Unsupervised person re-identification (RE-ID) has attracted increasing attention recently due to its low costs and high application values. Currently, most of the unsupervised RE-ID approaches attempt to explore discriminative visual features based on unlabeled samples, which suffer from the label noise problem. To alleviate this problem, this paper first proposes to leverage both visual and spatial information for unsupervised person RE-ID, with a feasible assumption that the same identity has similar neighbors across different cameras. Technically, we devise an iterative Graph Convolutional Network (GCN) to alternately improve the quality of pseudo-labels and multimodal fusion features. What is more, our modal is trained based on three novel pseudo-label classes (pseudo-positive, visually similar, and pseudo-negative) instead of two to better learn discriminative features. Extensive experiments on two representative datasets (Market-1501 and DukeMTMC-reID) demonstrate the effectiveness of our approach, with 1%-2% mAP improvements as compared to the advanced approaches.