How to stitch automatically panoramas of images that lack well defined elements, sharp edges?

Question

I often take large landscape panoramas of 100+ 21MPixel images. Quite often some of the images do not contain sharp, well defined elements that can be detected by automatic control point detectors. Typical examples of such cases are clear sky, clouds our water surfaces.

I use hugin on linux for the automatic control point generations. The quality of control points is often bad in images without well defined edges and the later stitching either takes very long as the optimization converges slowly or fails. Usually such images are placed in funny ways, even rotated.

Dropping such images is usually not an option as replacing these images with either black or white background ruins the rest of the panorama.

What could be a solution? Is there a way to put those images into their place? It is possible to move around images in the hugin preview windows, however the resolution is not good enough for accurate placing. It is also very tedious as severals 10s of images might have to be positioned.

I look for an automated solution if possible.

Answer 1

I was commenting on a similar question here: https://photo.stackexchange.com/a/26908/7736

-> Question link: Gigapixel panorama problems

The basic answer is, if you regularly create panorama images that contain not well defined features you will need a structured panorama, i.e. one in which the images are laid out in a structured grid, say 5x5 or 10x5 etc. If you do this once in a while, using a tripod with degree markings may work, I tried it once and failed actually. Otherwise, the "professional" solution to such panoramas is to buy a Gigapan robot. Given that you say that you regularly shoot such large panoramas, I think it may be a good idea to consider investing in one after some additional research on your side.

Having said that though, in my experience Photoshop or Microsoft ICE never had any issues on clouds in my panoramas but then neither is an option for you on Linux.

Answer 2

The answer is to stitch as much as you can automatically (by restricting the selection to images that contain enough well defined features to generate control points) and then fill in the gaps yourself using the water and sky images, simply arrange the images sensibly and blend the edges together (you shot with plenty of overlap, right?).

You may even be able to use the depth of field to judge whereabouts the water images go, this should give you an approximate vertical placement, you'll then have to decide where they go horizontally.

If it looks right then it doesn't matter if the images technically aren't in the right place.

For plain skies I would be tempted to just fill the area with an appropriate gradient. In the past when assembling panoramas I have snapped the sky as a single wide angle image and just upsampled and pasted that into the final panorama as it prevents any weird banding due to exposure shifts when stitching.