Flat colour pattern interpreted as relief

In several tests with painted structures on flat surfaces (thickness=0) I found embossed structures in the resulting mesh. Is there a setting or filter to correct that during reconstruction?
https://www.dropbox.com/s/ubrfbyfb2hwl79f/RC_embossed.jpg?dl=0

Hi fred.froehlich
How many images altogether for this example ? Reconstructed on normal or HIGH ?

In the Component were 270 of 277 pictures used. The screenshots from the first post were taken right after alignment with the default settings (Alignment mode: High). After processing > Reconstruction > High Detail the embossing effect looks even stronger:
https://www.dropbox.com/s/gpjyy9dtqafg6q6/RC_High_embossed.jpg?dl=0

I have had that too… mainly on flat surfaces with high contrast…

Hello,

what happen indeed is that the real surface of the solid black parts could not be recovered. This is not really surprising, because in general, photogrammetry methods need nice surface texture in order to work as it needs to visually identify the same points in more images. Finding which “black pixel in a spot” is which “black pixel” in an other spot is impossible even for human. In case of solid colors, there always remain at least one dimensional ambiguity, as geometrical constraints can be used too.
That means, there are more interpretations of the same 3D model (more models projecting to the same looking images as the provided ones). RC returned one of them.

What can help in such case is adding texture (e.g. using powder, projector, lamp,…)

What camera you use? Did you use RAW files or shoot in JPEG? Most of current camera processors use many technics (especially sharpening like) for improve quality of images or Jpegs.

Also probably you have not too much images. Small overlap can not allow reconstruct good quality depth and as result mesh can have errors or artifacts.

The black “blind” spot theory should be true for 100% black. Since the black areas are not interpreted as holes in the previous test I’m not sure if it applies here. I have done another challenge with 602 aligned RAW images also shot with a Canon EOS 450D:
https://www.dropbox.com/sh/ipem7rmb54uyk9h/AACjieqKZkoYvbb2JKWp3ApCa?dl=0
The vertical lines on the sails are actually threads with the same thickness as the others - they have just a different color (70-80% K versus 30-50% K). The same problem here: the generated mesh (High Detail) interprets the colored threads as elevation. If I want to use RC for scientific purposes - and let’s say scan the highly detailed surface of an oil painting - then that color error would really matter.
Is it possible to adjust a kind of threshold within RC or does it makes sense to filter the images before processing to avoid this negative effect?

Hi fred.froehlich
It is 100 % sure for FLAT surfaces, like the egg black stuff ( you can test it out on different FLAT and FEATURELESS objects, even though it is not 100% reproducible ) There is just not enough information for proper calculation. That is why in most cases are used some pattern projections to overcome this problem - the easiest and most elegant ways.

In case of the sails it is more interesting but looking at the model overall quality ( how many images ? ) it looks more like a sort of bad alignment… But that is just an idea what is the cause… there could be something different at play…
One of the images looks like the original photo, but even in a low-resolution downsample you can see a LOT of noise… what settings did you use to capture the model + what camera ?

The test with the sails consists of 602 aligned RAW images you can download a sample image from the dropbox link above.

Hi fred.froehlich
Yes, I have looked at the sample file and, as I have said, it looks too noisy… ISO 100 + 0,6 sec but it seems like a wrong setting = noise level too high… Is it possible to get the whole dataset in RAWs + RCPROJ files, so that we can take a look at this ?