On Wed, 20 Oct 2010 08:30:47 -0700 (PDT), kfj wrote: > > > On 20 Okt., 14:12, Robert Krawitz <r...@alum.mit.edu> wrote: > >> In general, fine tuning CPs in images offset at very sharp angles >> works very poorly; the rotation seems to confuse the fine tuning >> algorithm and the points are placed very strangely. >> > Did you try match_n_shift? It's a tool specifically to match images > with large distortions. It transforms the (e. g. fisheye) images into > something the SIFT algorithm can handle, then does the CP detection > and then remaps the CPs to the original image. Your problem matching > images from different orientations stems from > a) very different distortion of the ares you are trying to match > (match_n_shift may help) > b) parallactic errors, which are more of an issue than you may think > since your successive positions vary more likely by metres than by > centimetres, the way you describe you setting. > Also. different CPGs have different sensitivity to distortion. It > might be worth your while to try out a few to see which work best in > your situation.
That's not the issue. The original CP detection is working reasonably well, and this particular lens doesn't have particularly bad distortion in any case. The issue is the *fine tuning* algorithm. It's visibly finding the wrong place. Parallax is significant here, but that's not the issue I'm having with this. >> 2) Removing the CPs in the image pairs I didn't want to tie together >> had to be done one at a time; it would be handy to have a similar >> command to (1) remove all CPs between the selected images. > > If you go to the display of the control point list, you can sort the > list display by clicking on the column headers. This way you can > easliy get all the CPs for one image pair together, and you can just > mark and delete the lot without too much ado. That doesn't work too well. I can only sort on one column at a time. Furthermore, if I'm trying to delete all points between images 0 and 3, sometimes image 0 is the left image and sometimes it's the right image (so it takes at least two steps). >> 3) I'm mostly doing my own manual fine tuning by nudging each CP with >> the arrow keys. There are two problems here: > > You can aso drag and drop them. Sure, but that's a lot coarser (and more physically demanding -- holding down the mouse button while carefully moving the mouse) than nudging them with arrow keys. > You don't mention which version of hugin you use on what system, but > the newer hugin versions have masking of input images and the option > to remove CPs from masks. So you can also just mask whatever parts of > your images are giving you grief, have the CPs removed from those > parts, and take it from there. That would have been useful for another thing I had to do (remove all control points from something very close to the camera), but that's not the issue at hand.. I'm using Pre-Release 2010.3.0.14e86c2cfee4 on Linux. > Let me add a bit of advice: > I also thought that one needs a lot of control point with very good > finetuning to get a decent result. But I found out that other factors > matter more: > - for a very good result, you absolutely have to have good lens > calibration data. To obtain these, an image set with a lot of evenly > distributed and well finetuned CPs is needed, but once you have the > calibration data, you just need a handful of CPs for every image pair, > even two will do in a pickle. I'm sure it helps, but I managed to get a good result anyway. It's just that removing the excess control points and fine tuning the ones I wanted was much more grunt work than I wanted. > - if your data aren't parallax-free, like the data you're just working > on, no amount of fine-tuning will create a proper match between them. > I know that from doing lots of free-hand panoramas, and even though my > successive positions only vary by centimetres, it's an issue already. > Particularly if you have man-made structures (with their straight > lines and geometrical forms) in the images, there's hardly a way to > avoid some of the parallactic errors showing in the output, > necessitating manual intervention. Again, masking can help, since it > allows to make sure a certain structure is taken in it's entirety from > one source image and, hence, free from stitching artifacts. Also, > using a different stitcher (try smartblend) can help, though recent > versions of enblend also adress the issue if I'm not mistaken. I'm actually getting on the whole very good results. I have the worst case CP error down under 4, and only have a few visible seams. One of them doesn't matter (a marked line in a parking lot). Unfortunately, the other one does: the horizon of the ocean is off by about 1 pixel, which is quite visible. I've tried using horizon CP's, but that's still not entirely fixing it. That points out another problem: in the preview window, it's possible to click the horizon. Unfortunately, the preview cannot (that I have been able to determine) be magnified, so this too winds up being a coarse adjustment. Right now whatever I do I have the ocean somewhere tilting vertically by about 10 pixels. Not much, out of the 3456 I have on my 7D, about 1/4 of a degree -- but visible. On land, it would be irrelevant, and indeed on three sides of the shot (east, south, and west) there's enough land so that even though Cape Cod is fairly flat it has sufficient terrain to hide this. The open ocean to the north, however, can't. If the ocean isn't *perfectly* level, it's going to be noticeable. > - even if you don't put them in the final image, try and make a > zenith and nadir shot as well. They help consolidating the whole > show. If you are in a situation like the one you discribed, it also > helps to try and make as many shots as possible out of one position, > so stick the whole camera out and try and capture as much as > possible by tilting it this way and that. Then go to the opposite > side and just take shots mainly of that bit of the content you > couldn't reach from your first position, with enough overlap to your > first set, and use that to fill in the gap best as possible, > expecting the need for manual intervention. (I assume you've been on > a tower of sorts) That gives you two bad seems to work on rather > than four. You can try to have the seems, then, where there's the > least human-made content to ease the stitching difficulties. There was no possibility of a zenith or nadir shot. I was on the observation platform of a tower, with a ceiling and a floor. I tried taking shots from other angles as well, but I had a lot more trouble merging that then just the straight out shots (tilted down about 15 degrees) from the four sides of the tower. My lens was actually wide enough to get everything I really needed. -- You received this message because you are subscribed to the Google Groups "Hugin and other free panoramic software" group. A list of frequently asked questions is available at: http://wiki.panotools.org/Hugin_FAQ To post to this group, send email to hugin-ptx@googlegroups.com To unsubscribe from this group, send email to hugin-ptx+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/hugin-ptx
