Object-Oriented Cutout Data Augmentation for Tiny Object Detection

Abstract

With the rapid development of deep learning, generic object detection has been widely applied in many fields of real life. However, the detection of tiny objects is still a challenging task due to fewer features and limited information in computer vision research. To overcome this limitation, we propose cutout data augmentation aiming at tiny objects that are prone to occlusion problems and occupy only small pixel areas in the image. Precisely, we perform a cutout that combines the traditional cutout method of randomly applying a mask to the image with the method of applying a cutout by dividing a specific area of the GT box corresponding to the category with the largest portion and the smallest in size of the dataset. By combining both techniques, we improve the occlusion problem while the semantic information of tiny objects is intact, making it more robust. Overall, the experiments achieve great results in improving accuracy on the tiny object dataset, VisDrone2019 [1].