Unmanned Aerial Vehicles (UAVs) have recently gained popularity due to their simplicity and effectiveness in traffic monitoring and potential for rapid delivery, and rescue support. Moreover, UAVs have been employed as a supporting machine in data collection for object detection tasks, in particular vehicle detection tasks in object recognition. Although vehicle identification is a tough problem, many of its challenges have recently been overcome by two-stage approaches such as Faster R-CNN, one of the most successful vehicle detectors. However, many critical problems still remain, such as partial occlusion, object truncation, object multi-angle rotation, etc. In this paper, we combine the Generic RoI Extractor (GroIE) method with Dynamic R-CNN and Side-aware Boundary Localization (SABL) for both testing and evaluation on a challenging dataset XDUAV. Overall, 4344 images in the XDUAV dataset, divided into 3 subsets: 3485 training images, 869 testing images and 869 validating images were used. These consisted of six object classes: 33841 “car”; 2690 “bus”; 2848 “truck”; 173 “tanker”; 6656 “motor” and 2024 “bicycle”. With the ResNet-101 backbone, our approach showed competitive results compared with the original GRoIE method, surpassed by 1.2% on mAP score and by about 2% on most classes AP scores, except for the class 'tanker'.