A Curriculum Learning Framework to Boost Object Detection of Unmanned Aerial Vehicles

Accurate and efficient detection of objects in complex environments, especially the detection of aerial objects, remains a significant challenge in surveillance. Conventional object detection models often struggle with the intricacies of aerial imagery, such as varying sizes of the objects, distances, and environmental factors, leading to subpar performance. Additionally, simply using a state-of-the-art object detector tuned on aerial objects tends to perform poorly on smaller objects. This paper addresses these challenges by presenting a novel approach to enhance object detection accuracy and efficiency. Using a modified YOLOv7 architecture, we introduce a mixed curriculum learning (CL)-based framework tailored for object detection in aerial imagery. The core idea of CL, drawn from the phased learning process in human education, involves the strategic ordering of training data from simpler to more complex forms. In our methodology, the difficulty criterion is based on the size of the bounding boxes, with the hypothesis that smaller bounding boxes correlate with higher detection challenges. YOLOv7 is first pre-trained on a generalized object detection dataset, then our approach systematically uses CL to improve the performance on small aerial objects. Further, we show that our CL protocol can effectively learn from a combination of synthetic and real-world data, with the former addressing the concern with the scarcity of data when such data are challenging to collect. Our findings reveal that CL can provide a generic object detector with significant performance improvement. This improvement was particularly notable in detecting smaller and more distant objects.