3D pose estimation is a study of estimating human 3D joints from a single image, and it is widely used in industrial fields and applications. The performance of 3D pose estimation has dramatically improved with the deep learning. However, the lack of 3D data has always been a constant problem. To solve this issue, we propose multi-stage learning method that uses both 2D and 3D datasets. We achieved 92.0% accuracy with Human3.6M dataset and obtained natural 3D pose results on outdoor images.