Re-KISSME: A robust resampling scheme for distance metric learning in the presence of label noise.

Distance metric learning aims to learn a metric with the similarity of samples. However, the increasing scalability and complexity of dataset or complex application brings about inevitable label noise, which frustrates the distance metric learning. In this paper, we propose a resampling scheme robust to label noise, Re-KISSME, based on Keep It Simple and Straightforward Metric (KISSME) learning method. Specifically, we consider the data structure and the priors of labels as two resampling factors to correct the observed distribution. By introducing the true similarity as latent variable, these two factors are integrated into a maximum likelihood estimation model. As a result, Re-KISSME can reason the underlying similarity of each pair and reduce the influence of label noise to estimate the metric matrix. Our model is solved by iterative algorithm with low computational cost. With synthetic label noise, the experiments on UCI datasets and two application datasets of person re-identification confirm the effectiveness of our proposal.