I had some really great results and some awful ones. Almost all were better than a simple Poisson fill, which I did for
a comparison. Here are the steps I used to complete these images :
1. Obtain a border for the missing part, Then for each Image :
2. Resize the image to the same size as the source image.
3. Find the offset for the match image that causes the least discrepancy on the missing part's border.
4. Graph cut the match into the source with the missing part constrained to come from the match and the area outside the border (which is decently thick) constrained to come from the source image.
5. Poisson blend them together
6. Blend along shifted lines. I had a lot of images with seams created by the circular shift applied to the match image. So I did just a little blending along that line to make it stand out less. I probably should have done more. But I didn't have time to run it over for another 12 hours.
7.. Finally I assigned the image a success value based on the flow of the graph cut. It didn't seem like much else mattered. Since the graph cut really seemed to represent how well everything worked out.
The Best and the Worst
First My favorites:
This one has some sky at the bottom but looks good above that, it could be cropped
Here are some bad ones, You can tell cause the Poisson fills (on the right) are better.: