Affine Non-Negative Collaborative Representation For Deep Metric Learning

In this paper, we propose a deep metric learning method based on affine non-negative collaborative representation (DML-ANCR) for person and vehicle re-identification. Our method can adaptively generate a non-negative coefficient matrix for support samples per class and fit the query sample with the support samples in the affine subspace. We predict the query sampleƒ??s label via the residual between the query sample and optimal fitness. We formulate the affine non-negative collaborative representation learning as a meta-learning problem and present an episode-based approach to learning the best fitness to maximize generalization. Besides, we apply a hard mining strategy to improve the robustness of the metric. In experiments, we also introduce the re-ranking method. Results show our approach has achieved very competitive performance on the widely used person and vehicle re-identification datasets. It surpasses most baseline methods and state-of-the-art methods.