A comprehensive review of few-shot object detection on aerial imagery

Abstract

With the development of technology, drones, and satellites play an important role in human life. Related research problems receive great attention, especially in the computer vision community. Notably, the object detection models on aerial imagery take part in many applications in both civil and military domains. Although it has great potential and has achieved many achievements, it cannot be denied that object detection faces many challenges such as the small size and the quality of training datasets. The few-shot paradigm was explored to tackle that challenge. In this paper, we intensively investigate 55 state-of-the-art few-shot object detection methods using many different learning styles such as meta-learning and transfer learning. Moreover, we analyzed 12 aerial imagery datasets and benchmarked state-of-the-art methods on three popular datasets, namely, DIOR, NWPU VHR-10, and DOTA. These datasets reflect the richness of classes and the complexity of real-world conditions. From the experimental results and analysis, we discuss insights and pave the way to the future outlook of this research.