Stereo fusion

Here, visual inspection of the overlay suggested that the general disparity was very large (somewhere between 100 & 200 pixels on the original images!). Hence it was considered necessary first to work with a scale down of 100:1 (i.e., to produce resultant images of 24 x 18 pixels). When the pair of scaled down images were overlayed it was evident that there was the expected general level of disparity at the reduced scale, but also that the scene was very greatly simplified (to just a few main profiles).

The approximate disparity computed for this 100:1 scaled down pair of images was 1.01 pixels, so a shift of 100 pixels was applied to the second of the original images and then a 50:1 scaled down pair of images were generated. The approximate residual disparity computed for these new, roughly matched images was 1.128 pixels, or about an additional 56 pixels shift on the second of the originals. After applying such a secondary shift, a pair of 24:1 scaled down images was prepared, these in turn yielding an additional correction of 0.338 pixels (i.e., about eight additional pixels on the second original). Thus the total shift computed in just three simple steps was 164 pixels. When the two original images before correction were checked manually it was found that the approximate foreground disparity was 162 pixels while the disparity for distant objects was 168 pixels. So the computed correction for best whole scene fusion was about as optimal as one could expect for a 3D scene.

An animation of the various steps described above is shown to the right as a sequence of low resolution overlays interspersed with overlays of the originals with progressive corrections applied.