Announcement

Collapse
No announcement yet.

Stitching challenges NN5, Samyang 12mm, Sony A7rii

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Stitching challenges NN5, Samyang 12mm, Sony A7rii

    Hello John and others, long time no see ;)

    I've just completed a job and was not happy with stitching challenges in all of my 58 images taken... I struggled to make them acceptable, but it is frustrating given how little I charge and how much time it takes when things are not stitching easily and well.

    One of the mysteries is why PTGUI optimising does not happen (it seems) when you initially "align images". Sometimes I optimise having just created the panorama, and things shift and improve, and that happens even when I have not added or changed any control points/levels or anything at all. It seems to me that the same optimisation would/should have happened when it first aligning the images, no?

    My NN5 bracket is calibrated as well as I think it can be. I am using the From the three projects linked to below, the one set at 9.8(camera on arm) and 9.15(upright base)is the correct bracket setting but I was experimenting by moving one at a time 0.5mm at a time to see if there was a difference. Sometimes it seems that though the spindle on the Z axis (under the lens nodal point on the tripod) is perfect at 9.15, but there is a bigger stitching error on the floor, and sometimes when the spindle z axis is not a perfect circle and alignment is out, there are less errors on the floorboards of the room. The lens is set as a Samyang/Rokinon 12mm theoretical and 12.54mm focal length fisheye.

    Because I was frequently trying to optimise and kept getting "bad" or "not so good", I thought it was perhaps because there was not enough overlap for enough matching control points because I am shooting 5 around and one up... so I just tried 6 around and one up (also tried one down too for good measure but that made no positive difference)

    On the 6 around + 1 up, I see just as many, if not more stitching errors. After optimising, I get "The optimiser failed to achieve a good alignment of the images. Re initialise and try to optimise again?" every time. There is a huge overlap of images, almost 50% of each image is an overlap of the previous... so why/how could this be "not good"?

    Luckily this isn't my main source of income, but when I do get a project, it would be nice if I could more reliably get good results!

    Your input on these many issues would be appreciated,

    6 around one up https://drive.google.com/file/d/1zTk...ew?usp=sharing
    5 around one up https://drive.google.com/file/d/1M4y...ew?usp=sharing
    5 around one up https://drive.google.com/file/d/1iYJ...ew?usp=sharing


    Cheers,
    Mark.





  • #2
    Mark, I'm afraid that I have been unable to extract your files from the RAR archives. I have tried two Windows utilities (7-Zip and Universal Extractor), and both are unable to process the files. Can you generate zip archives please?

    John

    Comment


    • #3
      Ah! Third time lucky. I tried ZipPlus and that has worked ok. Files extracted.

      John

      Comment


      • #4
        Mark, I have obtained a good stitch from the third set of images. The project file can be downloaded from https://www.sendspace.com/file/6cnn2z . It's a PTGui V10 file, but can be loaded into V11.

        Your zenith image is upside down, I think. You probably have auto rotate set on the camera, which doesn't work sensibly when the camera points directly up or down. But you can tell by comparing what you would expect to see at the top and bottom of the zenith image with the overlapped images below. (See tutorial at: http://www.johnhpanos.com/upside_down_tut.html ). So my project file assumes you have rotated the zenith image by 180 before input to PTGui.
        I deleted all control points and generated points only between adjacent images 0-1, 1-2, 2-3 etc.
        All points on the tripod and panohead were deleted.
        No special treatment required for zenith image (Ie no individual lens parameters).
        All's well with the stitched output file (PTGui blender rather than Smartblend).

        I haven't looked at the other 2 projects, but if the above guidance doesn't help to sort out those in a similar fashion, let me know and I'll check those too.

        John

        Comment


        • #5
          I had a look at the second set of images. Project file is at https://www.sendspace.com/file/nhea1n . A problem in your project is the generation of control points on the panorama head. It helps if you mask the tripod and panohead from the camera images so that (a) you don't get control points generated there, and (b) those black areas don't get used in the blending.

          John

          Comment


          • #6
            EDITED!

            Hi John, my apologies for not responding sooner! I have two email accounts, and this one is for my 360 work, which I have not looked at for days. I though I was registered on this forum with my personal email... anyway, thanks for all the insight and time you've taken. I will check out your projects using my PTGUI v11... I have removed v10 from my computer.

            You're right, I did first rotate the zenith image 90 clockwise so it is vertical, the same as the others. I am just reading your tutorial saying it DOES matter whether I rotate CW or CCW. I have not seen any Auto Rotate setting on my camera. I don't think I can turn it off, so perhaps I should try CCW with my camera instead of CW as CW has been what I used for this shoot, and resulted in dodgy stitching.

            So the process you'd suggest is that I manually mask the pano head out of each image before generating control points, then once that's done, remove the masks to get it fully stitched? I see in the first project you sent, the zenith is about 180 rotated, despite having control points connecting to all the "around" images... so I'm a bit baffled there...
            It would be really nice if PTGUI would just refrain from creating points in that area! I will definitely give it a go.

            Thanks again John, you're a star.

            Mark.


            Originally posted by John Houghton View Post
            I had a look at the second set of images. Project file is at https://www.sendspace.com/file/nhea1n . A problem in your project is the generation of control points on the panorama head. It helps if you mask the tripod and panohead from the camera images so that (a) you don't get control points generated there, and (b) those black areas don't get used in the blending.

            John
            Last edited by markpalmos; 12-06-2017, 11:32 AM.

            Comment


            • #7
              Mark, I suggested masking the panohead from the images to prevent control points being generated there. But there is no reason to unmask those areas when you stitch. It is positively beneficial to retain the masking throughout. Why would you want to see it in the stitched image (apart from test shoots when checking head setup)?

              As for the camera's auto rotate feature. I looked in the manual and could find no mention of any way to switch it off. However, if you shoot RAW, you can check the orientation of any suspect images and rotate if necessary before converting. In the metadata (optionally displayed in Photoshop's Bridge), the Orientation parameter should be the same value for all the images (usually -90 when the camera is conventionally mounted on the right of the panohead). I think you had yours mounted on the left, so they should all be showing +90. In any event, ensure they all show the same value before converting the files. Whether all the images are rotated CW or all CCW is not important. It doesn't affect the stitching quality. And don't be confused by parameters yaw and roll on the Image Parameters tab for the zenith and nadir images. They are interchangeable, since rotating the zenith/nadir image can be looked upon as being a change in roll or yaw. Remember that the thumbnails on the Source Images tab always show in the orientation as opened by PTGui and will never change during the various stages of the stitching process.

              John

              Comment


              • #8
                Hi again...
                I loaded the first project DSC03623 9.8 + 9.2 V10.pts and it's much better, but still has quite substantial stitching zenith errors at the top left of the balustrade between 1-7 and 6-7. Also at the top left corner of the nearby door. The error in the floorboards going into the room with the near door is now gone, which is great... presumably from removing the control points on the pano head. Anything else I can try to get the zenith stitching better?
                tx
                Mark.

                Comment


                • #9
                  Hi John,

                  So for PTGUI to work at it's best, is it the actual orientation of the image or the consistency of what parameter it displays that matters? I ask because I would like to simply rotate the images CCW (since CW did not work well) and be done with it. I'd like to avoid extra steps like using Bridge (which I have avoided for many years due to terrible sluggishness). Presumably lightoom can tell which way it is, but is the displayed parameter different for an image rotated CW vs CCW, and if not, how would I know which way is the correct way?

                  WRT raw, On occasion I don't shoot raw as I use the camera's internal HDR for quick and dirty shooting.
                  WRT nadir, I actually like seeing the TOOL rather than one of those (IMO) ugly blurs or a black blob.

                  Thanks John,

                  Mark.

                  Comment


                  • #10
                    Originally posted by markpalmos View Post
                    Hi again...
                    I loaded the first project DSC03623 9.8 + 9.2 V10.pts and it's much better, but still has quite substantial stitching zenith errors at the top left of the balustrade between 1-7 and 6-7. Also at the top left corner of the nearby door.
                    Mark.
                    Mark, My stitch can be downloaded from: https://www.sendspace.com/file/vf7fe5 . I don't see any significant problems. One thing I noticed on one of the other stitches was a small degree of parallax visible at the bottom of the bannister where some hanging beads are seen through the gaps. It's not possible to adjust the panohead so as to eliminate all parallax completely for fisheye lenses, owing to the fact that the entrance pupil is not located at a fixed point but varies according to the angle at which the light rays enter the lens. So it's possible that some small glitches might be seen in some places.

                    The nadir in all my panoramas shows what is below the tripod, which I prefer to displaying the tripod or any other addition such as a mirror ball or logo disc.

                    John

                    Comment


                    • #11
                      Originally posted by markpalmos View Post
                      Hi John,

                      So for PTGUI to work at it's best, is it the actual orientation of the image or the consistency of what parameter it displays that matters?
                      Mark, Consistency is what matters. For simple PTGui processing, it is best to allow all the images to share the same global set of lens parameters. Included in this set is the horizontal angle of view (HFOV). All the images must therefore have the same HFOV, which means you must not have a mixture of landscape and portrait images. Further, the global parameters include the lens shift parameters, which enable PTGui to compensate for the lens axis not being precisely aligned with the centre of the image sensor. The shift parameters identify the size of this offset in pixels. If you were to rotate an image through 180 degrees, the corrections necessary to compensate for the lens offset would be in the opposite direction. Thus it's necessary to ensure consistency by not having some images upside down, which can happen with autorotate features on cameras not working reliably. The rule is that all the images shall be rotated identically (if at all) from the camera's natural landscape raw orientation. (Nothing to do with the angle the camera is physically held in). But it doesn't matter which of the 4 possible orientations you choose to present the images to PTGui in - they just have to be all in the same orientation. I hope that's understandable.

                      Of course, PTGui is well able to cope with any mixture of different lenses and an assortment of camera orientations, but this needs careful application of the individual lens parameters feature with much more complicated optimization, so its best avoided whenever possible.

                      John
                      Last edited by John Houghton; 12-06-2017, 04:20 PM.

                      Comment


                      • #12
                        Hi John,
                        This is odd, probably a v10 - v11 bug because when I open your DSC03623 9.8 + 9.2 V10.pts on my v11, it is not stitched nearly as well as the rectilinear image you shared here.
                        Running optimiser with your project shows "very good" but still does not get anywhere near the accuracy of the pano you shared...
                        So why would that be?
                        tx
                        Mark.


                        Originally posted by John Houghton View Post
                        Mark, My stitch can be downloaded from: https://www.sendspace.com/file/vf7fe5 . I don't see any significant problems. One thing I noticed on one of the other stitches was a small degree of parallax visible at the bottom of the bannister where some hanging beads are seen through the gaps. It's not possible to adjust the panohead so as to eliminate all parallax completely for fisheye lenses, owing to the fact that the entrance pupil is not located at a fixed point but varies according to the angle at which the light rays enter the lens. So it's possible that some small glitches might be seen in some places.

                        The nadir in all my panoramas shows what is below the tripod, which I prefer to displaying the tripod or any other addition such as a mirror ball or logo disc.

                        John

                        Comment


                        • #13
                          Originally posted by markpalmos View Post
                          Hi John,
                          This is odd, probably a v10 - v11 bug because when I open your DSC03623 9.8 + 9.2 V10.pts on my v11, it is not stitched nearly as well as the rectilinear image you shared here.
                          Mark, That is indeed odd. I have downloaded the project file
                          DSC03623 9.8 + 9.2 V10.pts using the link I posted and run it with PTGui V10 and V11. Both give an excellent stitch, but the optimization figures are slightly different:

                          V10: Average distance 1.76; Maximum 5.16
                          V11: Average distance 1.70; Maximum 5.95

                          The difference between our results may be due to different zenith images. Your original zenith needed to be turned upside down. I did a lossless jpeg rotation using the Windows utility Irfanview. Maybe your rotation produced a slightly different result for some reason. My rotated version is at
                          https://www.sendspace.com/file/91qo0x .

                          If necessary, I think it may now be more appropriate to pursue this offline via email.

                          John

                          Comment


                          • #14
                            Hi John, I'm on the train to London at the moment, back tonight... I'll check your rotated image against mine, and send you mine for good measure. 

                            I'm happy to move our conversation to email if you prefer. 

                            Speak soon, Mark. 
                             

                            Comment


                            • #15
                              Mark, I was right about there being a difference in our rotated image files, but the error is in my version! It seems that Irfanview's so-called lossless rotation option is not doing what it says, for the output image dimensions are different to the input file. The original file is 5304x7952, whereas the rotated version is 5296x7952. Having used Irfanview for more than 10 years, I find this very surprising and disappointing. So I rotated the image in Photoshop instead and made a new project file: https://www.sendspace.com/file/1nen44 . Not perfect, but you should find it better than the previous one.

                              John

                              Comment

                              Working...
                              X