Sample-Dependent Distance for 1: N Identification via Discriminative Feature Selection

We focus on 1: N identification tasks with binary features. Most multiclass classification methods, including identification and verification methods, use a shared metric space in which distances between samples are measured regardless of their identities. This is because dedicated metric spaces learned for each identity in the training set are of little use for the test set. In 1: N identification tasks, however, gallery samples contain rich information about the test domain. Given a sample and its neighbors in the gallery set, we propose a method for calculating a discriminative feature selection mask that is used as a sample-dependent distance metric. Experiments on several reidentification datasets show that the proposed method enhances the performance of state-of-the-art feature extractors.