Reconfigurable Hybrid Model Convolutional Stage – Infinity Laplacian Applied to Depth Completion

Convolutional networks are the current approach that presents the best performance in many applications. The principal critic of classical models is that they are hand-crafted by the designer. Still, the proposed architecture of a convolutional network is also hand-crafted, for example, the number of layers. Depth map completion is crucial for computer vision due to its applications in different fields such as video games or autonomous vehicles. Depth maps are acquired by a sensor or obtained by a stereo algorithm and present a lack of information due to occlusions or sensor misinterpretation. In this paper, we offer a reconfigurable hybrid model to interpolate depth maps. This model consists of a convolutional stage (SC1) pipeline, interpolation model, and convolutional stage (SC2). The convolutional stage input is a color reference image of the scene, creating a color features map as input for the next step. The interpolation model is the infinity Laplacian. We interpolated the incomplete depth map solving the Infinity Laplacian in a Manifold. Then, the completed depth map is processed again by the last convolutional stage. In this pipeline, we used a fixed number of convolutional filters, but we can interchange convolutional steps, i.e., interchange SC1 by SC2, reconfiguring the computing sequence. We estimated the parameters of the convolutional filter and the Infinity Laplacian using Particle Swarm Optimization (PSO). Our proposal obtained MSE=1.315 in the KITTI depth completion suite outperforming some contemporaneous methods.