Starter questions about Reconstruction

Comments

6 comments

  • Avatar
    Götz Echtenacher

    Hi Tom,

    1. The sparse cloud represents all the Tie Points, if they don't exceed the setting "Max points to display".

    2. That depends on what you want. You can adjust the settings for each step. Preview is by standard simply the sparse point cloud but you can change that. The differences between the 3 possibilities is mainly the resolution of the depth maps, which is the basis for the calculation of the mesh (or dense point cloud in other software).

    3. In Preview on standard settings yes. Normal not. What takes "so long" is the calculation. :-) Try other software and you will not complain again about the time it takes - RC is quite fast. The temp files are the depth maps.

    4. Geometry and texture are very different things. The texture is basically wallpaper glued to the mesh. So if you imagine a photo-wallpaper of a forest, it looks natural even though the wall is flat. The reason is that the texture shows detail that your brain interprets the way it is used to - natural shadows will enhance that. The mesh is unaffected by the texture.

    5. Yes there is, you need to click on the mesh in 1D tree view. Info panel shows specifics of the reconstruction region.

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Brilliant clarification - thanks.

    So, depth maps. This is RC's unusual approach, compared to the competition, which largely explains RC's speed?

    I'd really appreciate an explanation, or pointer to learned source, of what the two approaches are, and the pros and cons of RC's choice.

    'The temp files are the depth maps' - does this mean we mustn't delete them?

    Or, like the temp files of photos' features during Alignment, RC stores the result, so the features and depth maps can be deleted?

    During today, with both a good outdoor photo set and a problematic interior set, I have managed to more than double their Tris and Total projections but am shocked at how bubbly and lacking in detail and crisp edges, are the Normal Detail reconstructions. Sure I could get good building survey dimensions off them but not finer details of joinery sections etc.

    Will these miraculously clean up, with credible accuracy (not just smoothing-over) in later stages of the process?

    0
    Comment actions Permalink
  • Avatar
    chris

    its a bit better if you use high detail. and then drop the detail to something usable after.

    also taking higher res photos and getting closer with more photos helps too.

    but really rc and photogrammetry in general doesn't do a very good job with man made hard surfaces. at least not without going to extreme levels of detail to begin with.

    so for good results it requires more steps. either taking it to zbrush or re-modelling them in something else. and then taking those models back to rc for texturing.

    hopefully we get new algorithms in the future to improve this. along with a classification system for selecting different parts.

    0
    Comment actions Permalink
  • Avatar
    Götz Echtenacher

    Also, a laserscan can help enormously with plain walls, but I figure that is not what Tom is after...  :-)

    Depth maps are to my knowledge totally normal for that kind of software. They are, in my imagination as I haven't verified it with a simple search (hint-hint), just a grayscale image where the different hues represent different depth measurements.

    You can delete them, but then RC will have to re-create them if you want to reconstruct another part of the model. They are component specific which means they change if there is a different alignment. Image features are also stored only in the cache - only Tie Points (the 3D equivalents of 2D features) will be stored in the component since they form the sparce point cloud.

    In my experience, wobbly surfaces will not miraculously be flat in high detail as opposed to normal. The wobbles will just be much crisper.  ;-)

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Thanks chris.

    I'm beginning to see two different problems with photogrammetry on buildings, from an architectural POV - first with flat featureless surfaces, second with things like the linear moulded detail of joinery or carved stone.

    The first I understand - little or nothing to be grabbed as a distinctive and unique feature to locate in multiple views, as every bit of the surface looks same as any other.

    Beginning to realise that the second may be similar - though distinctively non-flat crosswise, they are still uniform linearly so nothing distinctive to locate lengthwise. I guess corners and curved work are easier.

    Both cases, at v high resolution, the system may find paintbrush irregularities, or natural grain if unpainted, but high res won't help with truly featureless factory finishes.

    So you suggest external re-modeling and I can see that would make it look nice - but reliable as factual accuracy? Linear mouldings would be idealised guesswork unless isolated bits came out clearly. Maybe flat surfaces could be inserted - but again relies on isolated bits of clear surrounding perimeter, to orient their plane. Do such vital clear bits actually happen and can they be relied on? Especially when it's window woodwork complicated by adjacent reflections.

    Improving algorithms - yes please! What do you mean by classification systems for selecting different parts?

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Gotz, thanks - answering in a minute

    0
    Comment actions Permalink

Please sign in to leave a comment.