Use of Synthetic Data on Object Detection Applications

Abstract

It is not always possible to find enough data in Deep Learning applications, where huge amounts of data is needed. One may face difficulties while trying to perform qualified works, where data cannot be collected in rare occuring events and situations involving privacy concerns. To overcome this issue, we create a military vehicle dataset in which great amounts of Synthetic Dataset can be found. Thus, with this dataset, we bring a solution to the problem of detecting military vehicles, which is an area where collecting sufficient data is a problem. In a training with 5, 10 and 50 epochs we raise the mAP score by 0.09, 0.03, and 0.02, respectively.