Anchored Kernel Metric Learning

This paper proposes the Anchored Kernel Metric Learning (AKML) method, which learns metrics by sparsely combining locally discriminative rank-1 basis matrices in the anchored kernel induced space. First, we apply k-means clustering to generate anchored samples, thus forming an anchored sample matrix. Based on the assumption that the basis matrices can be formulated as linear combinations of anchored samples, we then learn the combining weights by regularized LDA in anchored samples' neighborhoods. Finally, a logistic loss with nonnegative constraints is utilized to select and weigh these basis matrices for a PSD metric matrix. Under this framework, we incorporate multiple kernels into AKML for further boosting the performance. Experiments on two benchmark person re-identification datasets demonstrate that our approach clearly outperforms the state-of-the-art metric learning methods.