Arbitrary-oriented ship detection with RetinaNet embedded angle vector module

Abstract

As an important part of object detection, rotated object detection plays a great role in many visual applications involving ship detection in remote sensing images, oriented scene text spotting, and face detection etc. Since the different definitions of the angle and the periodicity of the numerical angle values, the angular boundary discontinuity will happen during the model training which decreases the detection accuracy. To address that, we propose an arbitrary-oriented ship detection with RetinaNet embedded angle vector module (RetinaNet-AVM), which adds a new angle vector module (AVM), improves the boxes localization loss function and applies a new boxes angle loss function. At first, the numerical angle values will be encoded into angle vectors by the AVM before data training. After the training, the angle vectors will be decoded into numerical angle values to generate the bounding boxes of the final results by the AVM. On the HRSC2016 dataset, RetinaNet-AVM could detect the arbitrary-oriented, large aspect ratio and densely arranged ships with the rotated bounding boxes, and the precision is 90.76%, the recall is 89.66% and the average precision (AP) is 86.71%. Compared with the RetinaNet-H, our model has improved accuracy, recall and AP by 4.88, 1.47, and 3.9, respectively. And compared with R2 CNN, RRPN and RRD, our model is still competitive.