LIU Ying, LIU Hong-yan, FAN Jiu-lun, GONG Yan-chao, LI Ying-hua, WANG Fu-ping, LU Jin
Object detection has been well studied based on the traditional manual features and deep learning algorithm. However, the research on small object detection has just begun in recent years and there are little research outcome available. Furthermore, most of the methods proposed are based on the traditional object detection algorithms with certain modifications so as to improve the accuracy of small object detection. Small object contains fewer pixels and has less features, and it is even harder to extract object features after down sampling. Hence, small object detection is a challenging task. Small object detection has a wide range of application requirements in the fields of automatic driving, remote sensing image detection, and criminal investigation. It has important practical value for the research of small object detection technology. In this paper, the existing research results of small object detection are summarized. Firstly, the existing algorithms are classified into one stage, two stages and multi-stages according to the number of stages for detection. The principles of RetinaNet、CornerNet-Lite、feature pyramid network (FPN) and other algorithms are described and compared. Secondly, this paper describes the application of small object detection technology in different fields, and summarizes the data sets such as MS COCO、PASCAL VOC、DOTA、KITTI and algorithm performance evaluation indicators. Finally, the challenges faced by small object detection are concluded, and the future research directions are prospected.