Turntable-Constrained Camera Pose Estimation

Norio Kosaka, Shinichi Higashino, Shuji Yamaguchi; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026, pp. 88-96

We study camera motion estimation for image sequences generated by single-axis rotation, such as turntable capture or orbiting cameras. In this setting, all views share a common rotation axis and differ primarily by a per-view rotation angle, forming a low-dimensional motion family. However, standard structure-from-motion (SfM) pipelines estimate unconstrained pairwise geometry and only afterwards attempt to fit an orbit, which can produce physically inconsistent trajectories when visual evidence is weak. We show that under single-axis motion, the essential matrix has a structured form in which translation is induced by the rotation of a shared orbit vector. This reveals a low-dimensional family of feasible essential matrices that couple geometry across views. Based on this observation, we investigate estimators that progressively enforce the turntable constraint, including axis-projected rotations, structured essential matrix estimation, and global orbit refinement. Experiments on synthetic sequences and a Blender-based turntable dataset show that incorporating the single-axis prior improves rotation estimation accuracy and yields more physically consistent camera trajectories.