Taking strips from different images when the width of the strips is more than one pixel works fine only without parallax. When parallax is involved, no single transformation can be found to represent the optical flow in the entire scene. As a result, a transformation that will align a close object will duplicate far objects, and on the other hand, a transformation that will align a far object will truncate closer objects. Also, rapid changes between aligning close and far objects might result in useless results.
In order to overcome parallax problems, instead of taking a wide strip (of N pixels), we can synthetically generate intermediate images, and use narrower strips. For example, we can take a collection of N synthesized in-between images, taking from each interpolated image a strip one pixel wide. In order to synthesize in-between views we can use various methods, such as optical flow interpolation [ 2 , 14 ], trilinear tensor methods [ 12 ], and others. In most cases even approximate methods will give good results for view interpolation.
The use of intermediate views for strips collection gives the effect of orthographic projection, which avoids parallax discontinuities. This strategy can be combined with the methods that were described in the previous sections as a preliminary stage, such that a complete solution is given for general motion in general scenes.
Adrian F Clark