Region Selection with Discrete Fireworks Algorithm for Person Re-identification

Abstract

AbstractVarious hand-crafted features with metric learning methods have improved the person re-identification (Re-ID) accuracy. Metric learning methods for person Re-ID mean to match the features acquired from different persons. However, not all information of the features is valid for metric learning. Compared to these metric learning methods, the region selection with discrete fireworks algorithm (RS-DFWA) is proposed in this paper for hand-crafted feature designing. RS-DFWA uses the fireworks algorithm after discretization to select the effective regions of the feature maps at the metric learning stage. RS-DFWA has a faster convergence speed and a better optimization accuracy so that the noise regions such as background features would be ignored. RS-DFWA optimizes the fitness of the discrete fireworks algorithm while training the deep networks for person feature learning. The method we proposed is validated on the CUHK03 dataset, region selection with discrete fireworks algorithm for the deep features achieve favorable accuracy. For example, on the CUHK03 dataset in single query mode, an improvement of mAP = +4.6% is obtained by RS-DFWA compared to the Baseline model.KeywordsDiscrete fireworks algorithmFeature selectionDeep learning