Serge's Rock


This page relates to the problem of constructing a good stereo pair from imperfect source images, using Panorama Tools.

References: see http://panorama.maandag.com/3drock/
as described at http://groups.yahoo.com/group/PanoTools/message/10162
and associated postings

A first successful strategy to make stereo pair:
Resulting PTGui project file: 3drockRik.pts

Residual error: rms 1.03 pixels with maximum vertical disparity on horizontal controls = 1.08 pixels (for image height 2082 pixels)

Resulting stereo pair (crossed eye).  Larger image



If allowed to optimize roll for both images, then rms 0.43 pixels with maximum vertical disparity = 0.96, but roll = 15 degrees.

In this case, much better results are achieved by locking the projection parameters for one image, and accepting a small amount of vertical disparity in the resulting fit.

What went wrong with the original shots?  Hard to say.  Looking closely at alignment between the rock and the background mountain, it seems clear that the camera was a bit lower for the left-eye image than for the right.  This tilts the camera-camera axis and may be responsible for the roll in "best fit". 

However all the errors are pretty small, and the system is sensitive to any remaining barrel distortion -- changing b from -0.020 to 0 changes the "best fit" roll from +15 degrees to -7 degrees.

For followup discussion on this problem, see http://groups.yahoo.com/group/PanoTools/message/10184 , which discusses rationale and an alternative way to construct this stereo pair.  An excerpt...

- turn off optimization for the "normal" control point
- turn off fov optimization
- lock first image yaw/pitch/roll = 0
- optimize second image yaw/pitch/roll/d/e

The resulting fit has maximum vertical disparity 0.92 pixels (vs 1.08
for the earlier approach with separate fov's), but it has about 210
pixels shift (15%) between the two images. Horizontal alignment is
easily finished in Photoshop, but requires a separate step, and I
doubt that one could see any difference in the final result.

Interesting problem...

--Rik Littlefield  (contact me by email or posting on the Panorama Tools forum)

Page last modified July 29, 2004.