Lightweigth Convolutional Neural Networks for Person Re-Identification

Abstract

Person Re-Identification (Re-ID) is an important task in video surveillance systems. Computer vision algorithms can be used to search, retrieve and localize the person of interest in a camera network. Person Re-ID research is an active research, and most of the researches use deep backbone networks, such as ResNet-50 and GoogleNet, which are complex networks with many parameters to train. However, it is computationally complex and time consuming to train these networks for Person Re-ID especially when it is lacked to have a good computational power. Therefore, effective lightweight networks are needed to perform Person Re-ID with low computational power capacity. In this paper, we evaluate and compare some lightweight networks which proved themselves in object recognition tasks. We compare their accuracies and complexities. Evaluation is conducted on a commonly used Market-1501 dataset.