Room Layout Estimation by Learning Depth Maps of Planes from 2D Layout Labels

Abstract

Room layout estimation is a basic mission in understanding the indoor scene and has a wide range of applications. Specifically, the task of room layout estimation is to segment an indoor RGB image with semantic surface labels, i.e., ceiling, floor, and walls. A straightforward solution for the problem is to estimate the geometry of the dominant indoor planes. However, since the ground truth depth maps of the layout is not easy to obtain, research in this direction is rare. In this paper, we focus on the cuboid rooms and propose an effective learning framework that can learn the depth maps of planes from 2D layout labels without ground truth depth maps. We employ a deep network to learn the surface parameters, which can be used to produce depth maps of planes and layout results. Then we propose stereo supervision mechanism that encourages the generated layout to be consistent with the ground truth layout segmentation along Z axis and layout edge on the image plane simultaneously, so that the learned surface parameters and layout results are reasonable. Experimental results show that our method can produce high-quality layout estimates and corresponding depth maps on benchmark datasets.