Geometric models are tools used to analyze geometric distortions in S3D viewing. For standard geometric models, 3D perception is predicted by calculating the intersection between the two projection lines from the left and the right eye to two corresponding onscreen points. However, standard geometric models fail in two conditions: (1) S3D images are captured by converging cameras (toe-in) and displayed on a flat screen; (2) the viewer’s head is rotated about the vertical axis (yaw) or depth axis (roll) relative to the screen. In these two conditions, vertical screen disparity with respect to the viewer is introduced. In this paper, we extend standard geometric models to cover the conditions with vertical screen disparity. To test our models, we conduct experiments using random-dot stereograms (RDS) with vertical screen disparities introduced by rotating RDS images. Experimental results show that the perceived depth is well predicted by the proposed model when the rotation angle is smaller than 20°. However, this paper introduces a retinal eccentricity model to analyze geometric distortions from the viewer’s perspective, which indicates that the geometric distortions only affect the binocular depth cue and not monocular depth cues. The monocular-binocular depth cue conflicts with motions may induce motion sickness in S3D viewing.