Expand to Detect: Enlarging Bounding Box Annotations for Small Object Detection

Detecting small objects, those with 32 × 32 pixels or smaller, remains a challenging task in aerial imagery due to limited resolution and contextual information. In this study, we propose enlarging bounding box annotations for small objects, thereby integrating additional shape and spatial context into the training process. With this approach, we seek to reinforce both the contextual information and leverage object shape characteristics during the training process. Experiments conducted on the VisDrone dataset demonstrate that using expanded annotations yields an absolute improvement of approximately 10% in mean average precision. Notably, detection accuracy for small object classes, such as pedestrians and motors, improves by over 20% in average precision, highlighting the benefit of including surrounding contextual information for small object detection from aerial images.