Region-aware Random Erasing

Object detection task, as a prevailing direction of computer vision, involves many challenges. One of its most general problems is overfitting. Random Erasing is a state-of-art Data Augmentation method for avoiding overfitting. However, it aims at classification task. When it is used to train object detection models, it sometimes discards the objects, then the bounding boxes correspond to some noise regions. To solve this shortcoming of Random Erasing, this paper proposes Range-aware Random Erasing data augment method. In training stage, Range-aware Random Erasing randomly occludes a part of foreground and a part of background rather than occludes a part of a whole image. By using this approach, we can not only enlarge our training dataset to reduce overfitting without discarding objects, but also reduce the impact of background information. By combing Region-aware Random Erasing with Tiny-YOLOv3 on two public datasets, Widerface and ILSVRC2015-VID, nice performance improvements in mAP are showed.