2500 images - Infinite process

Answered

Comments

8 comments

  • Avatar
    Lucia CR
    Manuel wrote:
    Hello,
    My name is Manuel Franquelo, im trying to process quite a complex project using Capturing Reality promo license:
    These are the specs of the computer im processing the data with:

    Windows 10
    Intel Xeon CPU E5 2687W v3 3.10 GHZ
    64GB RAM
    GPU Titan GTX


    Number of images: 2.500
    MPX: 50 per image (Canon 5DSR)
    Files: Jpge files

    The process is stuck infinitely at the "processing" alignment step 66% :(
    Relevant Alignment settings:

    Max features per mpx: 1.000
    Max features per image: 800.000 (With 50mpx its recognising around 574.000 points per image) (Is it completely useless to use that amount of points?)
    Image downscale factor: 1
    Max feature reprojection error: 1.5000 (Maybe I should lower this value down with the amount of points that are being recognized?)
    Detector sensitivity: Ultra
    Image overlap: High

    PD: My idea is to achieve the highest quality results as possible.
    Thank you,

    Kindest regards,
    Manuel Franquelo.

    I usually use these alignment settings:

    Max features per mpx: 10 000
    Max features per image: 40 000
    Image downscale factor: 1
    Max feature reprojection error: 1.000000 (recommended value 3px at most)
    Detector sensitivity: Medium
    Image overlap: Medium
    0
    Comment actions Permalink
  • Avatar
    Lucia CR
    This one looks similar to your issue: RC gets stuck and won't admit it
    0
    Comment actions Permalink
  • Avatar
    Wishgranter
    Hi Manuel Franquelo

    I will explain the alignment settings a bit more.

    Max features per mpx: 10k = it is preventing the software to detect too many points on a small area and therefore wrong calculation camera parameters.

    Max features per image: 800 000 (with 50mpx it recognizes around 574,000 points per image) (Is it completely useless to use that amount of points?)
    -> yes its USELESS and counterproductive to detect so many points, but it can be tricky to set a reasonable size...
    Based on my experience, I would use 80k as a maximum.

    Image downscale factor: 1
    -> this can stay as it is, use downscale only if the images are blurry

    Max feature reprojection error: 1.5000 (Maybe I should lower this value down with the amount of points that are being recognized?)
    -> leave it on default 2.0 pix

    Detector sensitivity: Ultra
    -> change it to MEDIUM as it detects even "unreliable" points and too much, so overall alignment reliability and precision can be "worse" as with Medium settings

    Image overlap: High
    Set it to MEDIUM. When using HIGH, then RC will use just the "central" part of an image that is the "best possible" regarding quality but it can get into alignment problems -> more COMPONENTS are created…

    PRESELECTOR
    -> 40k points

    from detected 80k points it will select 40k points used in actual alignment step
    0
    Comment actions Permalink
  • Avatar
    Manuel
    Thank you Wishgranter,

    I will try to process the 2500 img data chunk with those settings and report back.
    Won't having a lower feature reprojection error lead to a better mesh?
    The overlap is of 90% so using the central area of the image shouldn't be an issue right?
    Its great to understand preselector now :-) I guess that it will take the 40k best points out of 80k?

    Kindest regards,
    Manuel Franquelo.
    0
    Comment actions Permalink
  • Avatar
    Manuel
    I've just processed a 450 image data chunk taken with a medium format camera.

    Alignment Settings 1 on the left are the settings I've found to provide more points with less reprojection error.
    Aligment Settings 2 are your suggested settings :-) What do you think?

    Best,
    Manuel.

    Sin título.jpg
    0
    Comment actions Permalink
  • Avatar
    Wishgranter
    Hi Manuel

    if it is a MEDIUM-sized camera then the alignment settings need to be changed, as there are much more pixels in the image -> more detectable points.

    you detect 400k points but use ( preselector ) only 10k
    choosing from 400k points the 10k good can be tricky and not "good", it is better to use some reasonable numbers, like 80k points, and use half of the points. the reprojection error does not need to be ultra-low like other solutions as it does some tricky things when reconstructing. everything under or slightly over 0,5 pix is OK

    by reducing the detected points (400k ) you reduce even the alignment speed...
    0
    Comment actions Permalink
  • Avatar
    Manuel
    Its a MEDIUM format camera yes, and the other data set using a Canon 5DSR has equally about the same amount of pixels than the medium format camera. The point I was wondering about is: If I have a camera with a high amount of MPx (50) is it worth to bump up the alignment settings? Will it affect the final results?

    Your suggestion for the best possible results using a high mpx camera is to detect a maximum of 80k points and 1/2 of preselector (40k) and setting the reprojection error to 2.0 (which equals to 0,5pix?)

    Thank you!
    Best
    Manuel.
    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Wishgranter: "the reprojection error does not need to be ultra-low ... everything under or slightly over 0,5 pix is OK"

    Manuel: "... setting the reprojection error to 2.0 (which equals to 0,5pix?)"

    Is that correct - reprojection error is reciprocal of 'pix'?

    Or is the 0.5pix mentioned by Wishgranter actually the mean and median error in Alignment Report?

    0
    Comment actions Permalink

Please sign in to leave a comment.