Point clouds based 3D reconstruction plays a critical role in many robotics and computer vision applications such as robot navigation, autonomous driving and medical image processing. As precision is always the first concern of point clouds, in this thesis, we specify it into two concrete issues: how to remove clutters from a point cloud and how to construct it accurately. Consequently, a multi-view based RGB-D inpainting framework and a Bingham distribution based adaptive filter-type pose estimation algorithm are proposed. Compared with past work, our inpainting framework is able to deal RGB-D sequences rather than simple RGB ones and our pose estimation filter frees the users from manual tuning as required in the prototype. And experiments shows that each of them can outperform the state-of-the-art approaches.