Aiming at enhancing co-presence in multi-user augmented reality (AR) applications, we propose a novel pipeline to integrate real-time 3D face reconstruction technique into full-body avatar representation in the collaboration system. In the existing multi-user AR systems, the avatars are represented either by predefined 3D virtual characters with static face, or dynamic face with simple predefined animations. However, in applications where social presence plays an important role, e.g. in AR teleconference application, we prefer to see high fidelity 3D avatars and observe the facial expression of the other remote users during the conversation. Many methods are proposed to reconstruct 3D face from images, but they do not investigate how to integrate the reconstructed face into full-body avatars, and how to implement the model into real-time applications. In this work, we present a complete framework to generate high fidelity avatars of users, and efficiently update the facial expression of the avatars within a low latency. We implement the real-time 3D avatar reconstruction system to be applied in a AR teleconference scenario. Lastly we run experiments and a user study, which shows that our system can significantly enhance the social presence in the AR teleconference scenario and the real-time avatar update frequency is above 10 fps.