Robust Augmentations for Small Object Detection of Aerial Images

Abstract

Object detection is one of the most fundamental but important computer vision tasks. However, small object detection remains an unsolved challenge due to insufficient detailed appearances and additional noises. Meanwhile, aerial images and intelligent transportation systems are under the restriction of difficulties such as dense object arrangement, a large number of small objects, and different perspectives, compared with natural images. To deal with these problems, an adversarial-like data augmentation training is proposed in this paper to narrow the data gap between remote sensing images and natural ones. The difficulty of the remote sensing object detection is verified and analyzed firstly by the classic single-stage anchor-based detector RetinaNet. Then, the multi-scale and data augmentations are introduced to alleviate the mismatch between general detector training and aerial images based on the anchor-free state-of-the-art (SOTA) model FCOS. Experiments on the remote sensing dataset, NWPU VHR10, demonstrate the quasi-antagonism data augmentation method improves the both anchor-based and anchor-free SOTA detectors for natural images with significant margins and shows the effectiveness on aerial images, especially small objects.