On a 360-degree composite of my office, ICE will do an excellent job stitching the full-resolution images, but will inexplicably leave out one of the source images. This is the resulting image:
Note the missing section (black bar) in the middle.
If I reduce the resolution of the source images from 3072px wide to 1024 px, the stitching is atrocious but at least all thesource images are included. This is the resulting image:
If you copy both results to your local drive and use a viewer to flip back and forthbetween the two, you'll see all the artifacts in the second result.
In general, this is fairly typical of my experience with ICE: dropped images and/or wacky stitching. So far I have been using AutoPano and PanoTools Assembler with very solid results, albeit a more complicated process. ICE is simplicity itself to use and very fast, but quite unreliable.
Project file for full-res source images here. Project file using lo=res images here.
This is a pretty challenging sequence - 47 images, with some repeated structure. I know that it is frustrating that we don't provide any manual intervention for cases like this. We wanted to keep ICE simple - and as I mentioned before we don't intend to compete with AutoPano/PTGUI. ICE is state-of-the-art in how automated it is. However, with this set you are pushing the limits of the technology. I doubt that any other completely automatic program would come close to stitching this. Today the sweet spot for the ICE technology is a 2 to 10 image landscape or cityscape. If I look at the results being posted on flickr this is where our users are having the most success. For this example I would refer you back to Autopano or PtGUI. We'd love to say with certainty that this type of set could be automatically processed 100% of the time - but I don't think anyone can make that claim today.
That being said, you can get ICE to generate the full resolution result from the reduced resolution .spj file that you created (although you'd have to live with the misalignments that you are seeing). All of the parameters in the spj file are resolution independent so you can use a text editor to replace the image file names in the reduced res. spj with the filenames of the full-res images.
One general technique to giving hints to ICE is described in the following:
http://community.research.microsoft.com/forums/p/2204/3468.aspx#3468
http://community.research.microsoft.com/forums/p/3328/5476.aspx#5476
Essentially ICE works by automatically extracting and matching control points. Since ICE has no idea how the images might be aligned, it needs to extract control points everywhere in all of the images. If you mask out portions of the inputs that you know don't overlap then this can help the program along. Then again - doing this may not be any easier than the more manual programs.
Hope this helps.