Label-Noise Robust Person Re-Identification via Symmetric Learning

Person re-identification (Re-ID) methods based on deep learning has obtained impressive progress in recent years. And the training of the model usually relies on a large amount of accurate pedestrian identity annotations. However, label noise is unavoidable and has a negative impact on the model training process. But existing research pays little attention to Re-ID with noisy label. In this paper, we propose a label noise robust model for person Re-ID by symmetric learning, without additional clean labels or predefined label distributions. We symmetrically enhance cross-entropy loss through noise robust reverse cross entropy. The experiments demonstrated its effectiveness against various ratios of label noise on the three different benchmark datasets.