One of the applications of object detection and recognition are detecting object from aerial images. There are various further applications of object detection in aerial images, it can be used for tracking objects, implementation for unmanned aerial vehicles (UAV), traffic surveillance, and calculating objects. As the applications of detecting objects from aerial images get wider and more common, the accuracy of object detection becomes more important. This paper evaluates the effects of changing the backbone of YOLOv4 (You Only Look Once), the current state-of-the-art single-stage object detection method, with EfficientNet and EfficientNetv2, the image classification method. The experiment was done with the CARPK dataset, a dataset of aerial images of cars in a parking lot taken by drones that suit car detection. The result of changing the backbone of YOLOv4 with EfficientNetB3 manage to increase the detection accuracy to 99.23% in the CARPK dataset.