Autumn...

Autumn...

by Linus Kleen

submit your photo


Picture of the Week Themes
Suggest and vote on themes

Please participate in Meta
and help us grow.

Tell me more ×
Photography Stack Exchange is a question and answer site for professional, enthusiast and amateur photographers. It's 100% free, no registration required.
up vote 5 down vote favorite

I took hundreds of pictures from a 75mm-300mm lens in Cristo Rei(Lisbon), but since Hugin (panorama tools GUI) canoot distinguish well water from... water(neither i) i took the river pictures with a 18-55mm lens at 18mm(APS-C), and the remaining with 75mm (and some monuments with 300mm). My problem is that since the river-photos are much wider, and lower-res., hugin cannot find any matches nor know how to scale them.

To solve it, i tried to scale them (4 photos) by 376.77% (63.3º / 16.8º * 100) to match 75mm and put them as background, but each picture is so big that i have not enought RAM to process those huge pictures(I only have 3GB). How can i do it? Should i split every scaled photo into about 20 pictures? But then, how will hugin distinguish water from more water?

How do photographers usualy make rivers in gigapixel photos?

If you want to suggest an alternative GUI/program, then it must work in GNU/Linux.

share|improve this question
1  
For the movement, I asked a question after seeing a gigapixel crowd shot. Seems like an awful amount of manual work and still not sure it will always work! – Itai Sep 5 '12 at 15:37
1  
For memory limits, most computers now have more than 3 but not terabytes and that is OK. Good software is designed to handle this by splitting it up in smaller chunks and working with a few at a time. Now if you need to do this manually, it sounds too much to me. – Itai Sep 5 '12 at 15:40
2  
@Itai "A photo at 18mm is not just wider than 75mm, it has a different perspective" this is incorrect, perspective is determined by subject distance, so provided the camera didn't move you should be able to blow up the 18mm images and there would be no difference except for the resolution. – Matt Grum Sep 5 '12 at 16:09
1  
@Itai I'm not sure what you mean by coinciding at the focus plane but I am 100% sure that (ignoring DOF issues), if you rotate the camera there is no difference in appearance that cannot be corrected by an image transformation, i.e. if you rotate the camera right, you can warp the image so it looks like you rotated the camera left (except that your field of view is different). This is why rotating the camera about the centre of projection is so good for panoramas - the images will always line up. – Matt Grum Sep 5 '12 at 16:26
1  
@Claudiop There is absolutly NO need to have a lot of RAM for a Gigapixel panorama. What for? And only computing clusters can offer TB of RAM. What you need is a fast drive for temp files - i.e. as a scratch disk, a SSD is great here. A (cropped) 400MP panorama needed about 80GB of temp space with Photoshop CS4 on my laptop and took I think around 1 or 2 hours to create. – DetlevCM Sep 5 '12 at 16:27
show 7 more comments

2 Answers

up vote 3 down vote accepted

I got it to work. I had two trials.

First, I looked for a similar situation among my existing files, I found three from the same perspective at different focal lengths. They were much closer than your range, though (18mm to 55mm). Hugin could read the EXIF data and adjusted horizontal field of view (hfov, parameter v) accordingly.

Then I simulated your more extreme situation using an existing panorama and "photographed" it at different "focal lengths", i.e. I cropped out some details and scaled them down appropriately.

Here, Hugin choked. First no Exif data and thus no v-parameter to start off with. It could not find any matches. I then went and manually created control points. This is a bit of work (auto-adjustment won't work), and then had it optimized using position and view (view is important!). This worked surprisingly well.

Here is what to try: Load all images, have Hugin find correspondences. There will be two sets (ignoring 300mm for now) of connected images: The 75mm-set and the 18mm-set. Now, you should be able to manually create control points between images from the two sets. Then, optimise.

My Hugin version: 2011.4.0.cf9be9344356

Please let me know how this works out for you. If it doesn't work, maybe you could upload some pictures.

share|improve this answer
Oh... They're just about 200 images... It could be worst! :P Let's put hands working! – Claudiop Sep 6 '12 at 12:17
No, you won't need to do all by hand. A few should be enough to connect the two sets (total of say 10-20 control points). Then optimise, and then re-run finding correspondences! – Unapiedra Sep 6 '12 at 12:45
You solution seems to be working, but i can't ensure that it works yet. In the whole day i were only able to complete two optimizations, and there are huge improvements, but hugin is still unable to find control points between the 18 and the 75mm sets. This is a "minor problem" as manually adding everything just takes a couple of hours, the remaining time is spent processing... If this works, it should be complete within 48 hours. – Claudiop Sep 6 '12 at 19:49
1  
NOTE: It takes tons of time to render everything. It took more than 24 hours for every try. To speed up everything, is a good idea to make a 20Gb swap file. A big /tmp is also required, otherwise it will fail. Everything runs reasonably fast until you reach the end of RAM and SWAP space. – Claudiop Sep 11 '12 at 20:39
up vote 7 down vote

Question: How do photographers usualy make rivers in gigapixel photos?

Answer: They buy a Gigapan robot which shoots a structured panorama, i.e. a regular one in say a 5x5 grid and then use the software that came with their robot to merge the images. Having said that though, if you have such a structured layout, you could use Microsoft ICE just as well (as you can tell it how the images were shot). Here is an example of the SLR version of a Gigapan robot: http://gigapan.com/cms/shop/epic-pro (or maybe even two - I think a Dubai shot used two cameras on two bots)

Having said that, I managed to create an image that would be 400MP cropped (500MP uncropped with "blank spots) that included a river by just handholding in Germany. However I never had just water in the image but always some other background that enabled the software to stitch it (with only tiny errors). (used a 70mm lens that that time)

If you really want to regularly shoot Gigapixel images, I don't think there is a way around getting one of those machines. If you just want to shoot one every now and then, the best option is to buy a good old fashioned tripod head (like a Manfrotto 460MG) and then shoot an image say every 2 degrees, manually implementing the approach of the gigapan robot. Having said that though, you also want a lens with little vignetting, I tired doing that in Germany as well with a cheap old lens at 300mm and I just couldn't merge the images because of that... (and possibly not sharp enough), but if you have successfully created smaller panoramas, your lenses should be fine.

Wit respect to software: I personally use either Photoshop CS4 or Microsoft ICE with none coming out as a clear leader in terms of quality, however neither will work on Linux. I suspect you will find it a lot easier to find good software for Windows or potentially an apple product as these are more predominant amongst photographers than Linux.

share|improve this answer
BTW, I use both Photoshop CS5.5 and MS ICE under Linux (OpenSUSE 64bits) using Oracle VirtualBox. MS ICE actually has good image quality but had much better results from Autopano Giga which IS available on Linux, plus Windows and Mac. – Itai Sep 5 '12 at 16:38
The robots are interesting because they know with perfect precision where to stitch. Doing it by hand would introduce a margin of error which I am not sure would solve the problem. – Itai Sep 5 '12 at 16:39
@Itai Let's assume you do a robot image at 10x10 images and one by hand, you will end up with 100 photos. In either case you tell the software that it has to stitch a 10x10 grid - whether they are aligned perfectly or not does not matter. As long as you don't have any gaps in your images, the software will overlay them correctly - with the benefit of having an initial sort. If I can create a panorama by hand with 2 to 4 layers (vertically) and I don't know how many images horizontally and stitch those without any issues (assuming I have no gaps), then I cannot imagine any problems. – DetlevCM Sep 5 '12 at 20:18
No it actually matters. If the shots do not line up exactly, then the software has to match features and this requires a multi-resolution algorithm like SIFT. If they do line up exactly, then the stitch can be done with without regards to features first and then cleaned up via blending as an after phase. This has to play a big part when there are waves or crowd or anything that constantly moves. – Itai Sep 5 '12 at 23:03
And this still leaves the question as to what the problem is. So it needs to match features, so what? As far as I am aware any software to merge images does it anyway. Even a gigapan would need to do it - just in case the tripod moves in for example wind or the position isn't quite as exact due to for example the weight of the camera/lens combination. (And using a 300mm lens tiny differences would show) – DetlevCM Sep 6 '12 at 5:04
show 3 more comments

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.