Enhancing Defense Surveillance: Few-Shot Object Detection with Synthetically Generated Military Data

Abstract

Acquiring military-related data to train object detection algorithms for defense surveillance can be highly challenging due to security restrictions. To overcome this challenge, we utilize a few-shot object detection approach that can identify objects using a limited number of examples, deviating from the standard object detection methods that typically require large datasets for training. To compensate for the limited availability of military data, we employ generative models to create synthetic military datasets. This artificially generated data is then used as a support set to train the few-shot object detection network. We assess our method using a self-created dataset that includes four categories: soldiers, tanks, helicopters, and fighter planes.