Announcement

Collapse
No announcement yet.

Still struggling to create acceptable cameras

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

  • Still struggling to create acceptable cameras

    So I am still struggling to get the system to accept a full set of images.
    the camera I used this time was a 17mp Sony XZ1. known for its image quality

    I created a printed B/W "noise pixel" camoflage as a mat on which to photo the model - the idea being that at least the table top would provide anchor points in all the photos
    I then photographed a static object in the middle of this starting at the 4 cardinal points, then the 4 semi cardinal points and then the 8 points between the cardinal points
    this was done in 2 circular efforts - one with the whole 1.2mx0.6m table in view (object is about 10cm.) and a second round with the object taking up the whole image.

    and 3dFlow Light only identified 8 cameras out of the 38 that I had provided.
    image quality is high I will upload a sample of the photos when I get them cut down from the 2.4 mp that each photo is in jpg mode

  • #2
    Second photo

    Comment


    • #3
      First two photos were synched up into the mesh. the third one was excluded

      All three are very close in view.. and the actual resolution of the original images is roughly 4x what is uploaded and still the 3rd one didn't match

      they were run as unordered. even though they were done in a circular format.

      what am I missing

      Comment


      • #4
        Hi Karlsc -

        There are a few factors at play here:

        1. Black subjects are inherently tricky to capture, increase your exposure value to compensate for the dark surface and you'll see better results.
        2. Shiny subjects are also inherently tricky to capture because of the specularity shifting across the surface - polarization of your light source and the light coming into the camera will eliminate this issue.
        3. Your subject is taking up very little of the frame in two of the above images, especially the third. The reason the third didn't connect is that it's too far removed from the other too in terms of overlap - For best results try to keep the camera at the same distance from the subject, otherwise the features and level of detail change too much for a successful orientation to be possible.
        4. Instead of the pattern you currently have underneath the subject, I'd suggest using a single A3 sheet of printed perlin noise. Will have a higher density of unique features to aide orientation.

        What kind of camera settings are you using? This is a tricky subject with thin elements, dark surface and shine - You'll need to be shooting optimally to get great results. You'll most likely need a tripod and shutter release cable or 2s timer so as to minimise accidental camera shake, a cross polarized lighting setup would be ideal (or at the very least, lots of ambient light - outside, bright but cloudy still day) low iso, high fstop and fast shutter speed.

        There are lots of things to consider - I hope this is somewhat helpful, and if you have any specific questions, please fire away!

        Comment


        • #5
          No not really because as I described, the full set of images is two circular maps at roughly 18degress of arc separation. one at the longer distance to capture the background or random patterns on the table, and the other in close with the object taking up full frame, using fill flash. so the idea being that the 20 up close are overlapped but also given spatial context by the further out shots. and the further out shots given overlapping commonality points in the background textures.

          And from what I understand of how the algorithm works, that should work well. but it does not seem to.

          Comment


          • #6
            and now I crashed ZephyrLite :-P

            Comment


            • #7
              Hi again - I'll break down my answer into chunks:

              Roughly 18 degrees will provide too little overlap for a subject like this. You're going to need something more like 7.5 - 10 degrees for a greater chance of success. Most of the time with subjects similar to this I will end up with 288+ images otherwise I'm not happy with the results.

              Are you using a flash on top of your camera? This will be changing the specularity of the subject more than usual and probably causing big misalignment issues - you would be much better off with either xpola ringflash or outside with bright ambient light for this subject. A normal flash on top of the camera will work with more feature rich and easy subjects, but this is definitely not one of those.

              There is no point having the random patterns that far from the subject, they should be directly under the subject so that you don't have any sort of unnecessary distance disparity between feature points.

              Ps there are some fantastic resources to learn more about Samantha (the SFM stage) here:
              https://www.3dflow.net/resources/PhDThesis_RToldo.pdf &
              https://www.3dflow.net/wp-content/up.../02/cvpr10.pdf

              I highly recommend giving them a read to understand fully how the camera orientation stage works.

              For the other stages of processing, check out https://www.3dflow.net/technology/ for more deep reading

              Comment


              • #8
                Oh and if you had a crash, send the log to support@3dflow.net and we can have a look into what caused it and advise on how to avoid, or work on a fix if it's a software issue. Lots of the time hardware is to blame, but without the logs it's impossible to tell!

                Comment

                Working...
                X