Depth Completion using Convolution Stage, Infinity Laplacian, and Positive Definite Metric Operator

Deep maps have many applications, such as video games, autonomous driving, 3D reconstruction, and many others. Frequently acquired depth maps present holes or lack of information. Interpolation of the available data seems to be a suitable strategy to tackle this problem. Different methods have been proposed to solve this problem, from simple interpolation to deep learning. We proposed a hybrid interpolation model based on the Infinity Laplacian and convolutional stages in this work. The interpolator uses the available depth data and a color reference image of the scene to guide the diffusion process. Given the image domain and a proposed anisotropic metric, we embedded the depth data in a manifold, and the infinity Laplacian was solved in that manifold. The convolutional stage was created with Gabor filters of variable size. We also Pre-filter the available data to detect and eliminate outliers. Obtained results in the publicly available data set for depth completion show that the proposed model outperforms not only the previous version of this model but also similar state-of-the-art methods.