It seems you are asking quite a lot of things here. I'll try to provide you some information, but please notice: I'm not entirely sure of what you are asking here. Please provide me some more detail, and I'll do my best.
I'm trying to make an image of the
southern sky, horizon-to-zenith,
south-centered, on the iPhone. As many
of the images in the set will have
little detail, it's just sky,
image-matching stitching will not work
well. I'm looking for solutions.
What do you mean? Have you actually tried or are you just assuming so?
Following the basic outline here, is
it possible to skip steps 1 and 2, and
build the homography matrix directly?
I have the original alignments. It
would seem that would solve the
problem, as well as dramatically
reducing CPU which is an issue on my
iPhone.
Your question here is ill-posed. Step 1, that is detecting keypoints, serves the need to convert the image into a "feature space", that is to translate it in a coordinate map of points. You will later use these points to relate pair of images (in step 2).
You are doing this because you are trying to create a homography based on the matches you retrieved. A homography will serve you to reposition the pixels of the images onto a common surface. But, more importantly, you may say that the homography contains the information on the camera movements. If you already have information on camera movements, then you don't have to guess them from the images. Just feed the homography you already possess to the stitching pipeline.
If that is not possible, my next idea is to make large images containing multiple originals - if they always include the horizon at the bottom they should be able to match OK.
I don't understand what you are trying to say here.
So my main question (finally!)... does
OpenCV work OK with wide-angle images
like this? If I set my camera so I get
90 vertical I would get 60 degrees
horizontal, and with 20 degrees of
overlap (less, more?) that means about
five images to cover the 180 degree
horizon.
The more overlap area you have between images, the better, up to a certain point. 20% to 60% overlap seems reasonable to me, but it really depends on the materials you have, in the end. I'm working on 90%+ overlap area right now, just to give you an idea.
Now, at one point, you decide a common surface on which to project your images using the homography you computed pairwise-ly. In your case, I'd suggest you to chose a cylinder or a sphere as a surface. If you have troubles picturing this, imagine you stand in the center of a room with all of your pictures projected on the walls. What would be the best surface of the walls to be able to see the panorama without major distortions?
At this ... (more)
