Impossible alignment with repeated wall patterns?

Comments

72 comments

  • Avatar
    Tom Foster

    What rich info! Thanks

    0
    Comment actions Permalink
  • Avatar
    Götz Echtenacher

    Hey Steven,

    preselector to 250? Wow. Thanks for this idea. I think that was probably the most important one. Along with the CPs.

    Apart from that, you are making a great case about shaking it slightly until everything falls into place. At some point it's more of an intuition than exact sciense.

    I'm not so sure how much the white-balance helped. I'm pretty certain that feature detection can deal with different colors - features are more than that. I had no problems with aligning images that were taken at completely different times of the day and even with perfect white balance you could not avoid some changes in color. Same for images with and without flash (murmuring in the background) because in spite of tons of "advice" out there, flash usually works just fine. I think these tips stem from a time when the algorythms were still more error prone and they are also from many scientific sources, where people look more at the numbers than at the real-life results.

    Panos:

    Shooting several images from one position is not neccessarily impossible. In most cases it will align perfectly well, as long as the other requirements for photogrammetry are upheld. The problem CAN arise when RC picks 2 images from the same position to calculate the depth. RC will still come up with a slight difference for each position, so the base for such a pair is incredibly tiny, which will lead to excessive noise. So it's not that it won't work, but if there aren't enough images for RC to choose from, the quality can go down the tubes. The safest way to avoid trouble (I'm speculating here) is to reduce the overlap in the panos to almost nothing and use a tripod head that has its nodal point significantly off the sensor, so that with each turn the image position also shifts to an extent that is not "dangerous" any more.

    Best thread in ages!

    0
    Comment actions Permalink
  • Avatar
    Steven Smith

    Gotz I agree about the white balance and the sentiments about slight exposure changes. The software can handle a little, but just like changes in position need to be gradual so do these changes.

    I also a few idea about how feature detection works. I imagine it uses edge detection. To the RC it probably looks like this

    From low detection and few points

    to high detection with many detected features

    With maybe color Information overlayed Like

     

    And then adjusts the "perspective" or keystone of each image till features start to line up. It's probably smart about the process and starts with the pictures with the most similar features to speed up calculation. This is how I imagine in my head the software works.

    0
    Comment actions Permalink
  • Avatar
    Steven Smith

    here are all the files I used on this project

    https://drive.google.com/open?id=1_CZobR5gCslMfSCRJUKN9P26Luiwh5w8

    0
    Comment actions Permalink
  • Avatar
    Götz Echtenacher

    Hi Steven,

    nice images!  :-)

    Yes, I think you are right about the gradual changes. I haven't looked too closely at the image set.

    And it is also right that with a project that is beyond the limits it is a good idea to get as many variables as possible closer to ideal.

    0
    Comment actions Permalink
  • Avatar
    Steven Smith

    New project! take a bunch of pictures in terminator vision and see if it makes a model and find out what it looks like. There is a high chance it will be a waste of time and won't align, but if it does work it could be really cool ;)

    0
    Comment actions Permalink
  • Avatar
    Götz Echtenacher

    Hey Steven,

    very nice advice and examples! How does the model turn out? The angles between the shots might be a bit large and there should be more shots of the top, but I guess you've left those out...

    This st***d approval pending has thwarted me a few times as well. Nice workaround btw. The weird thing is that I get your notifications with the full text anyway!  lol

    0
    Comment actions Permalink
  • Avatar
    Steven Smith

    This small set is what i used to benchmark the new optane ssd i got. There is also a larger complete set for a different part of my testing.To be honest I didn't care so much about how the model turned out, just that between runs it was consistent.

    0
    Comment actions Permalink
  • Avatar
    Götz Echtenacher

    Totally get your point! Was just curious...

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    Nice work Steven ! Smart to have put down the preselector so low... We were having a discussion about how the preselection of features works on another thread, that's an interesting result !

    0
    Comment actions Permalink
  • Avatar
    Steven Smith

     I'm going to start a new thread. I'm having many of the same problems here on a much larger project. I will be talking about pano shooting and drone stuff if you guys want to join in.

    1
    Comment actions Permalink

Please sign in to leave a comment.