Workflow Q: control points and unused images

Answered

Comments

3 comments

  • Avatar
    Wishgranter
    Hi Benjamin von Cramon


    In most cases the alignment break in parts if there is not enough OVERLAP, so try to get more overlap in your capture next time.
    Delete all COMPONENTS except the largest one and use these settings for alignment..

    Alignment_settings.jpg
    0
    Comment actions Permalink
  • Avatar
    BenjvC
    Hello Wishgranter,

    Please, understand, I know I'm throwing a lot of questions here, and this is a forum, not technical support. II'll likely be purchasing a full RC license early next year, am using the present license to get me through developing several models for some strong prospects. The better prepared I am, the better for us both. Many thanks for your time.

    If you could clarify what in principle distinguishes a hopeless case of severely converged image pairs from those that the engine may not solve, but manual intervention with CPs would. Secondly, I'm unclear about the mechanics of that intervention.

    To the first question, here are a few steps and observations at each step, thanks for confirming my conclusion. I bring in 321 images, the largest component has 236. I inspect a corner of a building in this component and locate another component I know to relate to this corner, bring images from each component into 6 2D panels and set six CPs in as many images. My eye knows these features are legit, though admittedly the angle is pretty severe. I note high residuals for each these CPS, like 100, and see the red exclamation icon. Alignment brings little. Conclusion, overlap may be not the problem, but overly converged perspective is likely the culprit, there's nothing more to try in RC, go back an provide better photography in this area. Yes?

    Case two, I threw all the images of a different scene at one alignment effort, all cameras showed up in a single composition, alas I discovered ghosting upon reconstruction, double wall on a chair panel. I went back and began a second attempt, this time beginning with one group of images in this scene which had more in common, then incrementally added more groups of images and adding control points between certain views in one component to the next component. Eventually, all cameras are again stitched, but this time no ghosting, problem solved. So, here's an example where manual intervention can spell the difference, which then begs the question of an efficient workflow.

    If a scene splinters into components or presents misalignment that shows promise of intervention, then what's the best approach? I read that in working with components, one reconstructs one component at a time, exports, then imports the various components and uses control points to combine them. What exactly is being exported, the mesh? How would a mesh object provide what's needed to add control points to associated imagery? An FBX with cameras?

    Besides opening the poses folder under a component and closely studying what's in each to pick out suitable overlapping images for control points, is there a tool to make this search less manual? I'd love to select a component and have a window on a second monitor present a grid of thumbs for its group of poses, can this be configured presently?

    While simply adding more imagery to a single scene to grow a larger component may also get you there, my other concern is of exceeding the ideal mesh chunk size for use downstream. What workflow provides for nesting related components into a large scene, such that each constituent component can be reconstructed, exported, and later imported to downstream apps with a common scale and coordinate system?

    Again, many thanks for your time.
    Best,
    Benjy
    0
    Comment actions Permalink
  • Avatar
    Wishgranter
    Hi Benjamin

    Please, understand, I know I'm throwing a lot of questions here, and this is a forum, not technical support. II'll likely be purchasing a full RC license early next year, am using the present license to get me through developing several models for some strong prospects. The better prepared I am, the better for us both. Many thanks for your time.

    If you could clarify what in principle distinguishes a hopeless case of severely converged image pairs from those that the engine may not solve, but manual intervention with CPs would. Secondly, I'm unclear about the mechanics of that intervention.

    To the first question, here are a few steps and observations at each step, thanks for confirming my conclusion. I bring in 321 images, the largest component has 236. I inspect a corner of a building in this component and locate another component I know to relate to this corner, bring images from each component into 6 2D panels and set six CPs in as many images. My eye knows these features are legit, though admittedly the angle is pretty severe. I note high residuals for each these CPS, like 100, and see the red exclamation icon. Alignment brings little. Conclusion, overlap may be not the problem, but overly converged perspective is likely the culprit, there's nothing more to try in RC, go back an provide better photography in this area. Yes?

    YES... better images help not only with easier alignment but if they are properly taken, it can IMPROVE reconstruction results..

    Case two, I threw all the images of a different scene at one alignment effort, all cameras showed up in a single composition, alas I discovered ghosting upon reconstruction, double wall on a chair panel. I went back and began a second attempt, this time beginning with one group of images in this scene which had more in common, then incrementally added more groups of images and adding control points between certain views in one component to the next component. Eventually, all cameras are again stitched, but this time no ghosting, problem solved. So, here's an example where manual intervention can spell the difference, which then begs the question of an efficient workflow.

    CPs can help but they cannot do miracles, if the rules of photogrammetry are broken ( too sharp angles among images, panoramic shots, too big difference between close-ups and distant cameras - too different GSD )

    Besides opening the poses folder under a component and closely studying what's in each to pick out suitable overlapping images for control points, is there a tool to make this search less manual? I'd love to select a component and have a window on a second monitor present a grid of thumbs for its group of poses, can this be configured presently?

    use the POINT LASSO or POINT RECTANGLE tool and select features of the area which you see in your image(s), then click on FIND IMAGES to select the cameras that see that particular area..

    While simply adding more imagery to a single scene to grow a larger component may also get you there, my other concern is of exceeding the ideal mesh chunk size for use downstream. What workflow provides for nesting related components into a large scene, such that each constituent component can be reconstructed, exported, and later imported to downstream apps with a common scale and coordinate system?


    as mentioned earlier, with COMPONENT workflow you transfer ALIGNMENT results and as you add more of the components, they get together, and just then it is reasonable to do any reconstruction.
    0
    Comment actions Permalink

Please sign in to leave a comment.