In this paper, we propose a refinement filter for depth maps. The filter convolutes an image and a depth map with a cross-computed kernel. We call the filter a weighted cross bilateral filter. The main advantages of the proposed method are the filter fits outlines of objects in the depth map to silhouettes in the image, while the filter reduces Gaussian noise in other areas. Additionally, its computational cost is independent of depth ranges. Thus, we can obtain accurate depth maps at the lower cost than the conventional approaches, which require Markov random field-based optimization methods. Experimental results show that the accuracy of the depth map in edge areas goes up and its running time is low.