Centralized and Clustered Features for Person Re-Identification
Jian Lü, Yaozhen He, Tong Liu, Xu Chen · IEEE Signal Processing Letters · 2019
Extracting trusted label from unlabeled data and enhancing the discriminative ability of features are essential issues in person re-identification. The latest study shows that the progressive unsupervised learning has a good performance, and its advantage is specially shown in the ability to select reliable samples from the unlabeled dataset by clustering pedestrian features. Since the features extracted from the same pedestrian samples by the convolutional neural network trained by a softmax loss function may not have good clustering properties, a centralized clustering loss function (CCLF) is proposed in the letter. During the training process, CCLF makes the features of the same category close to the clustering center selected by $k$-means++, and implements it by adding a distance penalty term on softmax. In the evaluation stage, the study finds that the features have better clustering characteristics and the discrimination is enhanced. The experiments were evaluated on datasets CHUK03, Market1501, and DukeMTMC-reID to verify the superiority of CCLF.