Egocentric Indoor Localization From Coplanar Two-Line Room Layouts

Xiaowei Chen, Guoliang Fan; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 1549-1559

The coplanar two-line room layout with two parallel junction lines is often seen in an egocentric indoor vision when facing a wall or walking in a corridor. However, camera pose estimation from this kind of room layouts cannot be handled by existing vanishing point-based algorithms or PnL (Perspective-n-Line) methods due to the lack of line correspondences. This includes a recently proposed PnL-IOC approach that introduces image outer corners (IOCs), i.e., the intersecting points between room layout boundaries and image borders, to create more auxiliary lines. In this paper, a new coplanar P3L (CP3L) method is proposed to handle the coplanar two-line room layouts by embedding a P3L (Perspective-three-Line) method into the NSGA-II, a multi-objective optimization method. The proposed CP3L algorithm jointly estimates the initial camera pose and the 3D correspondence of four IOCs related to the two junction lines, and optimizes the camera pose in the iterative Gauss-Newton algorithm. We also study and compare the robustness of CP3L solutions under different configurations of auxiliary lines from estimated IOCs. Experiment results on both simulated images and real ones from the Matterport3D-Layout database demonstrate the accuracy and robustness of the proposed method.