The first 2 sets worked very well because there was no rotational invariance from
one shot to the next (as in, the shots were kept at a very similar incline- the
y value hardly changes at all). Features are matched very well and blending
looks nearly perfect even though just the very simple hat function is used.
What Didn't Work
The last set (the handheld pictures) obviously doesn't look as good as the 1st
2. I took these pictures with my own personal digital camera, and used Brett
Allen's method of camera calibration to get the focal length of my camera. I
think that when warping the images here, something went wrong. The focal
length was determined with a very primitive method (book, box, and ruler). And
I just had to estimate the k1 and k2 values based on what was on the webpage.
Considering all that guesswork, the panorama doesn't look half bad- features
are well matched.
Note: I manually cropped out the black top/bottom of .jpg images for
aesthetic purposes for the website, but this was not done automatically.
Cropping on the left and right sides were done as part of the program.