Accelerating Detection of Military Object in Remote Sensing

Abstract

This work introduces a dataset and synthesis tool for military target detection in satellite imagery using a virtual military simulation called ARMA 3. Additionally, it proposes the application of Dynamic Backbone Freezing to train the generated dataset, aiming to reduce training time while maintaining high performance. Currently, there are several real-world satellite image detection datasets available for academic use. However, there is no military target detection dataset for optical satellite images, and it is almost impossible to acquire real-world satellite images including military equipment for academic use due to cost and security issues. But, with the synthetic dataset generator proposed in this paper, it is possible to build high-quality, annotated datasets at little cost. We also check whether backbone freezing is valid in the field of remote sensing imagery detection for fast learning, and propose Dynamic Backbone Freezing that can significantly reduce learning time with maintaining high performance