The avatar-mediated 3D telepresence system aims to enable a user in local space to interact with a remote user through a virtual avatar representing the remote user. In this system, the avatar's movement should convey the meaning of the remote user's movement to the local space. If the local and remote spaces have different spatial configurations, the remote user's placement and motion should be adapted to the local space so as to preserve the motion semantics of the remote user. To this end, this paper presents a method to generate the placement and deictic gestures of the avatar. First, we develop a method to predict the probability of the avatar's placement in the local space according to the placement of the remote user. From this, we find the optimal placement of the avatar given the configuration of the local space and the remote user's placement. Second, we develop a simple yet effective method to retarget the remote user's deictic gesture to his/her avatar. User study shows that our method improves user engagement and social presence in the tested telepresence scenarios.