CVPR2021

Learning To Recover 3D Scene Shape From a Single Image

Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, Chunhua Shen

Abstract

Figure 1 : 3D scene structure distortion of projected point clouds. While the predicted depth map is correct, the 3D scene shape of the point cloud suffers from noticeable distortions due to an unknown depth shift and focal length (third column). Our method recovers these parameters using 3D point cloud networks. With recovered depth shift, the walls and bed edges become straight, but the overall scene is stretched (fourth column). Finally, with recovered focal length, an accurate 3D scene can be reconstructed (fifth column).