Conventional depth-camera can provide pixel-wise aligned depth and color images. However, the obtained depth image usually contains a lot of vacant image regions subject to the device resolution and reflectance property of target scene. To solve this problem, a novel depth image completion algorithm was investigated in this paper. To preserve sharp edges in the depth image, an edge mask was first designed. With reference to the edge mask, an improved joint bilateral filtering scheme was proposed. By filtering the depth image in four directions, a Markov random field model was used to combine the filtered depth images into one. Different from conventional filter-based image completion algorithms, the scene occlusion problem is also considered in the proposed algorithm. A variety of depth images are used in the experiment. Comparative results are presented to demonstrate the improvement over some classical methods.