Depth Completion Using Infinity Laplacian Based on Steering Positive Definite Metric Operator Plus Convolutional Stage

Abstract

Depth completion is an important task for autonomous vehicles, video games, and many others applications. Frequently depth maps present holes or data with low confidence level. This paper is devoted to the guided depth completion problem using a scene color reference image to guide the completion of a sparse depth map. In this paper we propose a model to complete the depth map based on an interpolator model and convolution stages enforcing color features of the reference image. In one hand, the interpolator we used is the infinity Laplacian (which is the simplest interpolator that holds a set of axioms). In the other hand, The convolutional stage is constructed using either Gabor filters (GF) or steering filters (SF). We tested different configuration of our proposal to find the best performance: i) SF Stage-infinity Laplacian-GF Stage. ii) SF Stage-Steering kernel-GF Stage. iii) SF Stage-infinity Laplacian (based on Steering Positive Definite Operator)-GF Stage. These models were assessed in the publicly available KITTI Depth Completion Suite. Best performance was obtained by the model iii) outperforming our model previous version, and others contemporaneous models.

Depth completion is an important task for autonomous vehicles, video games, and many others applications. Frequently depth maps present holes or data with low confidence level. This paper is devoted to the guided depth completion problem using a scene color reference image to guide the completion of a sparse depth map. In this paper we propose a model to complete the depth map based on an interpolator model and convolution stages enforcing color features of the reference image. In one hand, the interpolator we used is the infinity Laplacian (which is the simplest interpolator that holds a set of axioms). In the other hand, The convolutional stage is constructed using either Gabor filters (GF) or steering filters (SF). We tested different configuration of our proposal to find the best performance: i) SF Stage-infinity Laplacian-GF Stage. ii) SF Stage-Steering kernel-GF Stage. iii) SF Stage-infinity Laplacian (based on Steering Positive Definite Operator)-GF Stage. These models were assessed in the publicly available KITTI Depth Completion Suite. Best performance was obtained by the model iii) outperforming our model previous version, and others contemporaneous models.