Abstract: Depth images acquired by depth cameras generally contain noises and lose detailed geometric information. Thus, the filtering of depth images has become an important topic in both computer graphics and computer vision. However, most current filtering methods can hardly preserve the sharp features in the objects and often result in over-smoothing results. To this end, we proposed a novel joint bilateral filtering method for filtering depth images. First, we estimated the normal of each pixel in the depth image. Then we computed the weight of the normals by voting to perform joint bilateral filtering on all pixels. Finally, the vertex coordinates were updated according to the filtered normals. This method took into account the texture information with high accuracy as guidance information, which can yield more reliable filtering effects. In addition, this method was based on the local information of the point cloud, did not need to solve large matrixes, and employed GPU parallelism leading to extremely high computational efficiency. Experiments show that our method can highly preserve the edges in the normal field, thus preserving sharp features better than previous methods. 

Key words: depth images, point cloud, joint bilateral filtering, texture, normal filtering