Study On Rotation-Invariant In Aerial Object Detection Using Rotatable Bounding Boxes

Current methods in computer vision and object detection rely heavily on neural networks and deep learning.This active area of research is used in applications such as autonomous driving,aerial imaging,defense,and surveillance.State-of-the-art object detection methods rely on rectangular-shaped,horizontal/vertical bounding boxes drawn over an object to accurately localize its position.Objects in aerial images come in arbitrary orientation,in dense quantities and with complex surrounded making it a challenge to the traditional object detection models.And such orthogonal bounding boxes ignore object orientation,resulting in reduced object localization,and with overlapped boxes in dense situations limiting downstream tasks such as object understanding and detection.To overcome these limitations,this research presents object detection improvements that aid tighter and more precise detections.Firstly,we modify the object detection bounding box definition to include rotations along with height and width and the central point of the box to effectively locate the object more finely and reduce the interference of background pixels,afterword;we apply feature pyramid network(FPN)method to generate prior boxes from multiple layers to detect small-scale objects.Then,the regular encoding scheme for bounding boxes is modified to suit our rotatable bounding boxes for better estimation of the object orientation and boxes positions.To overcome the imbalance issue between the negative and positive samples,we combine hard negative mining(HNM)and focal loss(FL)to produces better results than using each one individually.We present results of object detection on four datasets and compared with(Single Shot Detector,SSD)and Faster R-CNN,showing that our proposed model surpasses traditional bounding boxes models,especially with rotated and dense objects.