Boosting video-based person re-identification with synthetic human agents

In recent years, research made in person re-identification has gained quite a bit of significance due to the increasing demand from a broad range of application fields with security and surveillance topping the list. A prominent part of this research utilizes deep learning methods that require large datasets with precisely extracted ground truth data. However, producing a large dataset from natural images for person re-identification poses many challenges. An alternative way of expanding the volume of available data is synthetically generating it. In this work, we present a synthetically generated dataset for video-based person re-identification that we created using real-world backgrounds and synthetically generated humanoids. Our dataset augments the DukeMTMC [12] dataset by simulating the scenes of the original dataset in our framework. Our dataset increases the size of the original dataset up to 3 times. This contribution improves the success rate of the Convolutional Neural Network based video based person re-identification approach by Wu et al. [34]. In addition to this, some tests conducted with the NVAN model of Liu et al. [23] to show that our method doesn't work in just one method, and we achieved similar achievements with this model as well. The results show that the improved dataset produced notably better results. Moreover, because of the generic format of our synthetic dataset generator framework, new datasets of different formats can be easily produced.