A depth map is fundamental for many computer vision applications, such as autonomous vehicles. A sensor like RealSense or LiDAR can acquire a depth map. However, depth data are incomplete due to sensor errors or occlusions between the objects in the scene. Completing this data in depth is a fundamental task for those applications. This manuscript shows an improvement of a depth interpolation model by adding morphological filters to the raw depth data to eliminate outliers and noise. This paper uses a model to interpolate data given by a variation of the infinity Laplacian. We have trained our model using three pictures of the publicly available KITTI database. The results show that our model plus morphological filter outperforms contemporary models. We empirically demonstrated that this kind of filter eliminates outliers and artifacts of the depth acquisition process.