Adaptive dense pyramid network for object detection in UAV imagery

Abstract

Object detection in Unmanned Aerial Vehicle (UAV) imagery has a wide variety of applications in both military and civilian fields. As UAV images are usually captured from flexible perspectives, with multiple altitudes, containing objects in various sizes and scales, such characteristics bring huge challenges to the object detection task. To address this issue, we develop a novel Adaptive Dense Pyramid Network (ADPN) that integrates congested scene analysis, aiming to incorporate the object distribution information into the object detection workflow without introducing additional annotations. We also design a creative Pyramid Density Module (PDM) for adaptive density prediction. The proposed ADPN incorporates the PDM and Object Detection Module (ODM) in a parallel manner to facilitate the feature alignment between density information and instance recognition. Our method can be applied to existing detection algorithms, and experimental results demonstrate the effectiveness and robustness of ADPN, and ADPN is able to improve detection accuracy on challenging datasets.