Graph Neural Network-Based Attribute Auxiliary Structured Grouping for Person Re-Identification
Geyu Tang, Xingyu Gao, Zhenyu Chen, Huicai Zhong · IEEE Access · 2021
Recently, person re-identification (re-ID) with weakly labeled or unlabeled data has drawn much attention in open-set and cross-domain re-ID systems especially for the attribute auxiliary weakly supervised person re-ID. Although state-of-the-art clustering-based methods have achieved good performance, the pseudo labels generated through clustering are often low-quality and noisy. To address this problem, we propose a graph neural network based Attribute Auxiliary structured Grouping (A2G) to improve the confidence of the pseudo labels. Different from the existing clustering-based approaches that only utilize the similarity in feature space, we learn the feature representation from the similarities in both attribute space and feature space by graph learning on the pedestrian attribute graph. Specifically, we first utilize the pair-wise attribute similarity to represent the relation between two pedestrians to construct a pedestrian attribute graph. Furthermore, we aggregate the features from their neighborhood on a pedestrian attribute graph by the graph neural network, which would make the attribute similar pairs closer and simultaneously take the dissimilar pairs further in the feature space. Finally, to avoid the over-confidence of the hard pseudo labels, we regularize the learning of the embedding model with the smoothed pseudo label (SPL) in the optimization of the feature embedding network. We conduct extensive experiments on several person re-ID datasets to validate the efficacy of our proposed method. The results demonstrate that our technique is effective and promising for person re-ID tasks.