Pedestrian Re-identification - Metric Learning using Symmetric Ensembles of Categories

Abstract

This paper presents a method for pedestrian re-identification, with two novel contributions. Firstly, each element in the target population is classified into one of n categories, using the expected accuracy of the re-identification estimate for this element. A metric for each category is separately trained using a standard (Local Fisher) method. To process a test set, each element is classified into one of the categories, and the corresponding metric is selected and used. The second contribution is the proposal to use a symmetrised distance measure. A standard procedure is to learn a metric using one set as the probe and the other set as the gallery. This paper generalises that procedure by reversing the labels to learn a different metric, and uses a linear (symmetrised) combination of the two. This can be applied in cases for which there are two distinct sets of observations, i.e. from two cameras, e.g. VIPER. Using this publicly available dataset, it is demonstrated how these contributions result in improved re-identification performance.