The intent of this paper is to introduce a novel method for 3D vehicle pose estimation, a critical component of augmented reality. The proposed method is able to recover the location of a specific object from a single image by combining pretrained reliable semantic segmentation and improved single image depth estimation. Our method exploits a novel pose estimation technique by generating new 2D images created from the projections of rotated point clouds. The rotation of the specific object is able to be predicted. Augmented objects can be shifted, rotated and scaled correctly based on the recovered orientation and localization. Through accurate vehicle pose estimation, virtual vehicles are able to be augmented accurately in place of real vehicles. The effectiveness of our method is verified by comparison with other recent pose estimation methods on the challenging KITTI 3D benchmark. Further experiments on the Cityscapes dataset also demonstrates good robustness in the method. Without requiring ground truth 3D vehicle pose labels for training, our model is able to produce competitive and robust performance in 3D vehicle pose estimation.