Comparation of Rotational Deep Learning based Methods

Unmanned Aerial Vehicles (UAV) imagery-based object detection has become more and more widespread, especially in fields like agriculture, city planning, construction oversight or even military reconnaissance. This research focused on a review of the state-of-the-art object detection algorithms which use oriented bounding boxes. The images were collected using DJI Tello and Zll SG906 Pro 2. The images were annotated using Computer Vision Annotation Tool (CVAT). The different detectors were trained using MMDetection and MMRotate modules from MMLab framework. The advantages and disadvantages of different deep neural network approaches are aimed to be compared: Rotated Faster R-CNN (two-stage anchor based detector), Single Shot Alignment Network (S
²
ANet) (one-stage anchor detector) and FCOSR (one-stage anchor-free detector). On neural network models two series of measurements were performed. In the first version, all the image samples were trained during the epochs. In the second version the images were grouped in tasks, and in each epoch a new tasks were load to train. Based on the measurement results, it can be concluded that better results were achieved with the Rotated Faster R-CNN model and the S
²
ANet model, the FCOSR performance lagged behind the previous models.