Panorama Fun
liebling at cs.washington.edu
Skip to Project Information
Sample Image
Kaidan Head Image
Handheld Image
Project Information
I took three sets of images; two outdoors in Redmond and one indoors in my office, although I only stitched the first two set (plus the sample set). I was amazed at how well the stich worked on the handheld images even though the alignments varied by ± 100 px horizontally and ± 40 px vertically. Of course, the color balance is not quite right but there was no initial color compensation.
The Lucas-Kanade algorithm seems sensitive to initial conditions. For example, in both the sample solution and my own code, if I set the initial estimate to be > 50 pixels away from the actual global minimum, then it would settle into a local minimum. One way to help this is to run with more levels in the Gaussian pyramid. For the handheld panorama, I used four levels and three iterations; for the Kaidan head images I used just three levels and two iterations per level. For the handheld shot, I manually aligned the images to see what reasonable estimates were to seed the LK algorithm.
My cylindrical warp is not quite right (i.e. it seems to be concave rather than convex projection; could be caused/fixed by a sign change of the focal length; the distortion works properly though). Otherwise the Lucas-Kanade algorithm runs within a few hundreths of a pixel of the sample solution -- probably due to some calculation differences. Again I have to say I'm very impressed with the LK algorithm; it's simple and powerful.
February 9, 2005
