Confidence-Aware Learning for Person Re-identification with Noisy Labels

Accurately labeling a large amount of data for person re-identification is a significant challenge. In this paper, we introduce a technique to effectively perform person re-identification even in the dataset with noisy labels. Leveraging the widely observed phenomenon that data with wrong labels tends to have large loss values, we fit the Gaussian mixture model(GMM) to estimate confidence which is the probability of the sample being noise-labeled. We propose confidence-aware learning that appropriately reflects confidence to balance between mitigating the impact of samples with noisy labels and guiding anchors to the complete positive and negative samples. Additionally, we refine the GMM to enhance the accuracy of confidence for each data sample even in a lack of data situation. Experimental results demonstrate that our methods are effective techniques for handling noisy labels in person re-identification.