Linear Quasi-Parallax SfM Using Laterally-Placed Eyes

Frontally placed eyes versus laterally placed eyes: computational comparison of their functions for ego-motion estimation

Both frontally placed eyes and laterally placed eyes are popular in nature, and although which one is better could be one of the most intuitive questions to ask, it could also be the hardest question to answer. Their most obvious difference is that, at least as supposed in the computer vision community, stereopsis plays the central role in the visual system composed of frontally placed eyes (or cameras); however, it is not available in the lateral configuration due to the lack of overlap between the visual fields. As a result, researchers have adopted completely different approaches to model the two configurations and developed computational mimics of them to address various vision problems. Recently, the advent of novel quasi-parallax conception unifies the ego-motion estimation procedure of these two eye configurations into the same framework and makes systematic comparison feasible. In this paper, we intend to establish the computational superiority of eye topography from the perspective of ego-motion estimation. Specifically, quasi-parallax is applied to fuse motion cues from individual cameras at an early stage, at the pixel level, and to recover the translation and rotation separately with high accuracy and efficiency without the need of feature matching. Furthermore, its applicability on the extended sideways arrangements is studied successfully to make our comparison more general and insightful. Extensive experiments on both synthetic and real data have been done, and the computational superiority of the lateral configuration is verified.

Estimating relative camera motion from the antipodal-epipolar constraint

This paper introduces a novel antipodal-epipolar constraint on relative camera motion. By using antipodal points, which are available in large Field-of-View cameras, the translational and rotational motions of a camera are geometrically decoupled, allowing them to be separately estimated as two problems in smaller dimensions. We present a new formulation based on discrete camera motions, which works over a larger range of motions compared to previous differential techniques using antipodal points. The use of our constraints is demonstrated with two robust and practical algorithms, one based on RANSAC and the other based on Hough-like voting. As an application of the motion decoupling property, we also present a new structure-from-motion algorithm that does not require explicitly estimating rotation (it uses only the translation found with our methods). Finally, experiments involving simulations and real image sequences will demonstrate that our algorithms perform accurately and robustly, with some advantages over the state-of-the-art.