Enhanced Single Shot Detector for Small Object Detection in Drone-Capture Scenarios

Abstract Harnessing the remarkable advancements in deep learning, object detection has witnessed a significant leap in performance in recent times. Nevertheless, the detection of small objects remains an intricate task, primarily due to their low resolution and noisy representation, particularly in drone-captured scenarios. In this paper, we introduce an Enhanced Single Shot Detector (ESSD), tailored for the detection of small objects in drone-captured imagery. Acknowledging that larger objects are typically easier to detect than smaller ones, we employ a scale-confusion erasing module to mitigate the noise introduced by larger objects during the detection of smaller ones. Additionally, we devise a neighbor fusion module that effectively integrates semantic information from deeper layers into shallower ones. Our extensive experiments and evaluations on the VisDrone2019-DET benchmark dataset clearly demonstrate the proficiency of ESSD in detecting small objects, exhibiting comparable performance to state-of-the-art approaches.