Automatic Generation of Stereo Images from Multiple Monocular Object Views

Abstract

Traditionally, to produce binocular stereo images, each single object view requires two shots with a proper camera displacement to generate a stereo pair. The total number of images required is twice as much as for monocular views. Special rigs, such as a slider, are also required for camera displacement if a normal single-eye camera is used. So, it is of advantage if the stereo views can be generated from multiple monocular images automatically. In this report, we present an image-based approach generating the stereo images automatically for multiple monocular object views. This approach does not include or require 3-D object reconstruction. The image reprojection technique is adopted and ensures the resulting images are physically valid. For fully calibrated images, the stereo pair can be generated directly using the image reprojection while for the uncalibrated case a preprocess is required to obtain the relative camera projection matrices. The concept of epipolar geometry and its role played in the perspective camera projection matrix recovery processes are then discussed in depth. If the viewing angle between two camera positions associated with two object reference images is sufficiently large, the closer in-between views are synthesized based on 3-D morphing.