I’ve mainly been working with natural materials, stone, just now working different types of projects in RC. I’m noticing a bumpy quality in reconstructions of surfaces that in reality are really smooth. I’m attaching some screen captures from Maya of a model from RC of a table. I’m also attaching crops of downresed source photos from similar angles to show how smooth most of the surfaces are, but also to show that the cause isn’t likely the photography, which all looks this way. Any thoughts on what’s causing this bubbling of tris on smooth surfaces and what might be done to alleviate it are much appreciated.
Every camera has a certain amount of noise. That is what you see here.
If I take lossless screenshots of a virtual 3d environment and reconstruct those, plane faces are perfectly plane.
Ah, that makes good sense. Your test with reconstruction of lossless screenshots was really clever. I’ll spend more time pushing what’s possible between denoising the source imagery and optimizing for ISO, not to invite the noise in to begin with.
Thanks!
Noise yes, but I think in this case the wood grain is also influencing your result.
If you look closely, there are significant areas with the same hue.
Also, it seems like there is some wood with iridescent grain.
Plus the reflective varnish doesnt help.
Considering all that, I think your results aren’t bad at all.
How detailed do you need the model to be?
I think if you want smoother results, you also need to go closer with your camera - so that you get images where basically every pixel has a different color…
ShadowTrail, you also get thumbs up for your test! 
All very useful information. I’m actually fairly close, considering I’m shooting with the 42 MP Sony A7Rii/21 mm Zeiss/ at about 1-2 meters at most. That said, at 100% I can see if I moved closer there’s yet way more detail to record in the grain, which then will allow me to also back away from ISO 800 and the accompanying grain. I’m cross polarizing, seeing minimal specular reflections, the 4+ stop light loss being why I’m having to push ISO so high, that plus stopping down to f8 and f11 when getting in close not to invite soft pixels on converged planes. I’ll try all this, add Nik denoising and post results.
I can smooth polys, of course, but the more procedural, the better. How good does the model need to be? This is an experiment to see what’s possible. Will then study the cost factors to know what quality level can be marketed accordingly.
Hi Benjamin von Cramon
use SIMPLIFY and get it down to say 10 to 100k tris. and see how it improve noisiness and sharpness of the model.
I simplified from 30 M to 5 M and see that the macro shapes in the bumps persist. I had yet to play with the Smoothing tool, am playing with that now. I’ve done plenty of manual smoothing with a brush outside of RC, which works great, but looking for procedural workflow naturally.
Thanks.
Wow, that is quite some serious shooting!
Is that for a museum presentation?
I read somewhere not to worry too much about soft edges - I think it might even have been Wishgranter.
So I guess rather not stop down too much to gain on ISO…
I am working with museums, but mainly for natural environment content. There I’ve played with resolution in the geometry and textures for A/B comparisons in Unreal, have to say you can’t have enough resolution of either. I’ve currently loaded 41 M poly model of a section of cave into UE4 with dozens of 4K textures (will step up to 16K with Granite for UE4 - texture streaming), and the rock keeps looking more and more like the real thing with higher poly count. You see the sharpness really speak with the camera in motion (motion parallax) and also in 3D.
As for tables and chairs and manmade environments, these obviously tend to be rich in smooth (easy to make) shapes, need the sharpness in the joints, but smooth surfaces want to be without bumps. I think Wishgranter said something about background pixels being soft helping separate a sharp foreground object during reconstruction, if that’s what you’re thinking of. I can’t see how mushy data is ever what you’d want, not even as a workaround to artifacts invited by grain or problematic subject matter. Clearly, never a one size=fits-all solution. We do agree, grain from ISO is baaad. The fidelity of a reconstruction of fine furniture isn’t something I’m after for any particular end goal, other than to raise the bar on a largely procedural workflow for virtualized objects and environments supporting mutable lighting. Still cutting my teeth ;^)
Hehe, well phrazed! (the penultimate sentence)
I think he said something like not to worry about parts of the object being out of focus if you have enough on other images to make up for it. Apparently RC filters the blurry parts or rather gives them less weight. But that’s only hearsay at this point. I’m sure the maestro will say something sooner or later…
Oh and I found that another way of avoiding bumpy areas (which can also be caused by small misalignments, although not in your case) is to “overshoot” the object, meaning you shoot, say, 4 - no better 10 - times the resolution you are after in your result and then just smooth it down or reduce polycount. For example your table would be magnificent for a rendering of a whole room - no need for more detail. With the texture on it, nobody would miss the small detail…
Hi Benjamin von Cramon
Read it one more time. im writing about 10-100k tris not milions… try it out please 
use SIMPLIFY and get it down to say 10 to 100k tris. and see how it improve noisiness and sharpness of the model.
Oh, didn’t realize that. An algorithm to bias sharp pixels, I like. I’m aware that Simplify uses adaptive subsampling to apply greater decimation to smooth areas than to rough, so why not go further with this? If a surface is known to be (nearly)planar, or cylindrical, conical, spherical, etc. smooth, why not have adaptive smoothing to apply as well? Also, regarding how grain from ISO figures in to a reconstruction, you’d think that it being random between images that the photogrammetry engine would easily filter out grain from the solution. Thanks, Wishgranter, for your insights.
About getting rid of noise: just reduce the resolution by half (quarter of overall number) and there should be no visible noise left. I once noticed this with an older camera with quite some noise - after reducing, it the images were brilliant. You would loose resolution but theoretically gain accuracy. And with 42 megapixels it should still be ok, I guess. Worth a try?
I can see situations, like the wood table’s top surface, where you don’t care about resolution in the planar geometry, but do care about resolution in the wood grain texture, might point to downscaling photos for reconstruction, then replacing the images with the original full res versions (same file names) for texturing, would zap the effect of grain. I’m still unclear why grain would factor in to begin with, given it’s random. Perhaps, it’s because the bumpiness is also random and based on however RC looks at grain and mistakenly determines certain grains per feature to be a common point of interest or feature. Generally, the actual features determine the basic shape of the table, then the grain creates a bunch of smaller scale errors that fall within a certain range.
Hi Benjamin von Cramon
Can you send me the data for close up inspection on my side ? its in RAWs or ?
Thanks, Wishgranter. At full res, which is the only way to really know what I’m dealing with, this is a whopping upload. I appreciate your offer, will possibly still take you up on it, but I’m running the imagery through Nik D’fine2 denoising to see what happens. At ISO 800 on the Sony A7Rii, you’d think the grain wouldn’t be so bad, but it’s pretty bad. I’ll post an update with comparison imagery of the reconstruction. Thanks.