EFPNet: Effective Fusion Pyramid Network for Tiny Person Detection in UAV Images

Abstract

Unmanned Aerial Vehicles (UAVs) have found extensive applications in the field of rescue and navigation scenarios. The objects in UAV images are generally with small sizes, which rises a serious challenge of object detection. Most existing methods address this issue by constructing multi-scale feature pyramids to integrate deep semantic information with shallow layer, but these networks fail to effectively extract and learn features of tiny objects in the shallow layer. In this paper, we propose an Effective Fusion Pyramid Network (EFPNet) for tiny person detection in UAV images. EFPNet consists of a Multi-Dimensional Attention Module (MDAM) and an Effective Feature Fusion Module (EFFM). The MDAM learns the weighted combination of features in both channel and spatial dimensions, which generates attention maps. It enriches semantic information in features. The EFFM utilizes the information from attention maps of different layers, which guides feature fusion between adjacent layers. It maintains consistency between deep and shallow features. Our proposed model achieves an Average Precision (AP) of 60.72% on the TinyPerson dataset, which demonstrate our model outperforms other state-of-the-art detectors.