Decoupled Teacher for Semi-Supervised Drone Detection

Abstract

Aerial drones come in handy in a variety of science and research applications, sometimes even cause privacy harms during aerial surveillance. In many anti-drone situations, detection, tracking, and classification of drones are of great significance to secure the airspace. Drones are small in size, different in appearance, and have different flight attitudes for various flight environments, which makes the drone data sets too expensive to annotate and the foreground-background imbalance occurs in drone detection. Previous works on drone detection have focused on supervised learning, which depends on large labeled data set. To alleviate the problem of scarce labeled data sets, Semi-Supervised Learning can be employed to leverage unlabeled samples. In this paper, we propose a Semi-Supervised Object Detection method Decoupled Teacher to use unlabeled data and address the foreground-background imbalance issue. Specifically, Decoupled Teacher decouples the Exponential Moving Average mechanism in the general SSOD paradigm, and applies a fusion method of weak/strong data augmentation. We have bench-marked our method and the state-of-the-art SSOD methods using two publicly available drone data sets. The experiment results demonstrate the superior performance of our approach compared with the state-of-the-art methods.