A Novel Plane Prior Assisted Depth Estimation of Indoor Scenes

Depth estimation of indoor scenes is still a very challenging problem, because these scenarios generally have many large, low-textured areas. Since these regions often have strong planar properties, many methods use the plane models to assist to solve the problem of depth estimation. However, existing plane prior generation algorithms do not fully exploit the plane information in the image, and the solution for solving plane information is not robust. In this paper, we propose an innovative scheme to get a planar prior for indoor scenes. Firstly, a reliable points selection method is developed. Then, due to the powerful performance of deep learning in plane detection, we use it to detect plane regions in images. We utilize the reliable points to find the best fit parameters for each planar region and optimize the plane models specifically for low-textured regions. Finally, our novel planar prior are embedded into the PatchMatch MVS pipeline to get the final depth map. Experiments demonstrate that our method can effectively recover the depth estimation in untextured areas.