In traditional stereoscopic displays, the virtual three-dimensional object does not appear to be fixed in space as the viwer's head moves. This apparent motion results from the fact that a correct stero iamge can only be formed for a particular viewpoint and interpupillary distance. At other viewpoints, our brain interprets the stereo image as a slightly skewed and rotated version of the original. When moving the head, this skewing of the image is perceived as apparent motion of the object.
This apparent motion anomaly can be overcome with head tracking. Unfortunately, a head-tracked stereo-display system requires the generation of images from arbitrary viewpoints. This has previously limited the practical use of head-tracked stereo to synthetic imagery. We describe a stereoscopic display system which requires the broadcaset of only a stereo pair and sparse correspondence information, yet allows for the generation of the arbitrary views required for head-tracked stereo.
Our proposed method begins with a pair of hyper-stereo reference images. From these, a sparse set of corresponding points is extracted. Next, we use image warping and compositing techniques to synthesize new views based on the user's current head position. We show that under a reasonable set of constraints, this method can be used to generate stereo images from a wide range of viewpoints. This technique has several advantages over previous methods. It does not require an explicit geometric description, and thus, avoids the difficult depth-from-stereo problem. We also describe a unique visibility solution which allows the synthesized images to maintain their proper depth relationships without appealing to an underlying geometric description.