Dense depth completion based on piecewise planar model

A new method is proposed to predict a complete and dense depth map from very sparse depth measurements. Previous state-of-the-art methods tackle the problem mainly by regular sparse depth distribution and exploiting the learning-based framework, which cannot guarantee effectiveness in various scenes. To handle these drawbacks, we propose a piecewise planar model based method, which models the depth map and corresponding color images as a collection of 3D planar, then transforms the task to the optimization of the planar parameters with the energy minimization formulation. Thus, the depth values can be computed through the fitting of the planar. The method can preserve the boundaries well and get high quality visible dense depth maps. In addition, our method doesn't need the training phase thus could robustly work on any scenarios, even if the sparse depth samples are irregular distributed. Apparently, the proposed new method has two applications: depth completion for Lidar sensors and converting irregular sparse depth samples computed from Simultaneous Localization and Mapping (SLAM) to dense depth maps. The method has been tested on the KITTI dataset and achieved the competitive results, proving its effectiveness in the real problem.