This sequence clearly illustrates some types of features are handled well by Lucas-Kanade alignment, and some which aren't. The buildings, with many unique details and solid, well-defined edges, are generally handled quite nicely. The fine repeating brick pattern on the ground is not; notice in particular that the diagonal concrete seams on the ground (which do not contrast much with the surrounding brick to begin with, and are partly obscured by the water) are badly misaligned in several places.
Parts of this sequence probably look better than they "should" because the large expanses of fairly regular water and sky do not need to be perfectly aligned in order to appear reasonable. There are some places where misalignment of certain features is clearly a result of differences in camera orientation between shots (because other features in the same shots are correctly aligned). Interestingly, we took another panorama from the same location using the tripod (not shown), and while the structures in the foregound align much more cleanly, on the whole it did not come out as well as this one.
If you view this in applet there is an obvious seam as you reach the edge of the image; this is not a problem with my blending function or skew correction (note the red square panorama is fine), but a result of some modest attempts to touch up the image in the GIMP after it was generated. I ran a sharpen filter on the image, and that the two edges should come out different is actually, on reflection, fairly obvious; it was probably a convolution filter, and if they didn't wrap the image cylindrically when filtering, the pixels on opposite edges would be filtered with completely different support and come out looking differently, even though they came from the same neighborhood of the same original image.
