Proposal (added 9/10/2008)
I would like to propose a project that involves combining pieces from one (or more) photographs into one 'base' image, based on optimally lining up hard edges between the photo piece and the base image. For a more complex example, consider the following:
Base image:
second image:
Piece taken from second image:
Oriented and matched to edges in the base image:
Final image:
A more simple example might involve taking two photos, cutting them in half, and then lining up the two halves by scaling/translating one half to best preserve hard edges (or other criteria such as matching colors, gradients, etc).
There might be a few interesting ways of looking at blending the boundaries of the two images in either case. Some sort of graph cut and/or poisson blending to both preserve strong details while blending over flat regions could help to eliminate the obviousness of the shape of the boundary of the image piece.
Other (unrelated) proposal ideas:
Photo-fractals
The idea here would be to start with one photo, allow the user to select a small sub-region of the photo. Based on some descriptor of the sub-region, do a search in low res versions of other photos and find a new photo that best matches the pattern of the sub-region of the first image. Do some sort of blending to insert the hi-res new photo into the sub-region of the first image allowing the user to 'zoom in' to the new detail (which would be completely unrelated, different scale, etc but would match the overall global pattern). This process could be repeated for an infinitly 'zoomable' photo.
HDR from a single photo
Another idea that would use a large image database, this would start with a photo for which there were parts that are underexposed or washed out in places. We could then do a search in our database for images that look similar either in the local neighborhood of the washed out area or just from the overall 'look' of that region. Then we could either cut and paste parts from other photos to fill in details, or perhaps better yet, do texture synthesis using the similar region in the database photo as the source image to fill in the details of the region in the original photo.
