Ripple / staircase artifacts on 3D reconstruction

[atorgon]

Hello

I am trying to use PhotoScan Pro 0.9.1 to reconstruct a 3D scene from 2 camera stereoscopic rig. In most of the reconstructed scenes, I get waves or stairs on surfaces which are actually flat. I have noticed it happens especially in regions with rather homogeneous texture. Does someone know why does it happen or how can I change the configuration of the program to correct that effect? I have already tried the option "Mild" when creating the mesh and it does not help.

You'll see attached the left and right images and the 3D reconstruction obtained from them.

Thank you

Alex

[atorgon]

Sorry, I forgot the images

[Wishgranter]

what settings use for MESH reconstruction ? it is on PointCloud clearly to see too ?

[James]

Looks like you are doing some 'undistortion' on your input images (curved black margins suggest anyway).

Is that to convert from fisheye to rectilinear?

Try the latest pre-release version 1.1.0 to process the raw fisheye images.

[James]

Perhaps also the blockiness/pixelated/step effect of the black margin correlates with the steps in your model?

[atorgon]

Hello and thank you for your replies!

To Wishgranter:
The wave effect is not apparent in the point cloud. There is some noise, but no periodical features. The waves appear in the dense cloud. I have tried mild vs aggressive filtering with little effect. On the other hand, the effect is more visible when I build the dense cloud in low quality than when I do it in high quality.

To James:
The images that I have posted are indeed distorsion corrected, but from a perspective camera, not fisheye. But the effect appears also when using images with no distorsion correction (and so, with no jagged edges).

Any other hint to solve the problem would be very helpful...

[Alexey Pasumansky]

Hello Alex,

To my mind the main problem is very small stereobase compared to the distance to the surface being reconstructed. Cameras are looking almost parallel and actually see the same scene, so the uncertainty of the distance from camera to the tie-points is high.

[atorgon]

Hello Alexey and thank you for your reply

The stereobase is 10 cm, distance to the target around 2m and focal length 12mm. According to my calculations, that gives a disparity of 0.6mm in the sensor, which is more than 40 times the size of the pixels. In those conditions, our own reconstruction software (in this particular applicationwe use Photoscan as a backup/reference) works fine...

I have tried to lock the calibration parameters of the cameras (stereobase, focal and distorsion coefficients), as those are well known. However, it is very difficult to set a stereobase of 10 cm fixing the GPS locations of the cameras! Maybe there is a more clever way to do it? I imagine this application is a little bit out of the intended domain for Photoscan (I think you have a product Stereoscan or something like this better suited), but it would be very helpful if we could manage to use it at least as a backup solution...

[Pascal]

Hello,
I continue this old topic, because I am not sure it is really closed ...
As discussed before, ripples in dense clouds may result from stereo base but may be not only.
In the attached file you can see two application cases, a complex and a very simple.
In the second one, the stereo base is about 20cm for a subject distance of about 1m,
and very similar to the others cameras.
So a too small stereo base seems not to be the correct explanation.
This issue is very critical in some application, notably when you want to use Photoscan to digitize
very fine engraving on stone for example (in archaeological research) because the risk of having false information is very important.
Can this problem come from the bit depth of the depth map? filtering or not filtering these maps do not change the result.
I am curious from the real explanation of this issue and I hope there is a solution in a near future. Others photogrammetric solutions based on PMVS do not produce such artefact, so a solution certainly exists.
Thank. Pascal

[Drakhain]

Hello,

I also have the same issue and do not understand why it happens.
I'm using a precalculated calibration obtained with agisoft lens, and let photoscan ajust them during alignment.
I always end up with ripples on flat surfaces and i do not know what to do.
Can someone help me?

Thank you.

[bigben]

Looks like a similar effect in panoramic stitching software from using a coarse (but quick) interpolator for transforming images.  Not sure if this would apply in this case, but I'm assuming there would have to be some transformation of the image to identify matching patterns between images(?)

I've seen something like this in some of my projects, typically near the outer edges of images, or where there are only 2 overlapping images.  Some of this may have been related to the optimised lens parameters (I was using fisheye lenses and optimising the parameters for the outer areas of the images is prone to variation) Getting better values for the lens parameters produced cleaner meshes in relation to this.  If you get good values for a camera/lens combo I wouldn't optimise them for a project, at least not for the initial alignment.

[Drakhain]

Hi,
Yes, it seems that optimization in combination with sparse cloud filtering does that, especially when i use a precalibrated camera for my chunks,but it do not understand why.

Recently i set tie points limit to. 0 allowing me to have enought points to filter.
I noticed that i get cleaner dense cloud, and it seems that having less reprojection error speed up dense cloud generation (all chunks's images still aligned)

But sometimes ripples appear, how can i correct this?
Should i let photoscan determine camera calibration during alignment, or be less brutal with filtering?

[bigben]

If you have lens settings that provide you with clean results then don't optimise them later (or at least don't optimise k1, k2, k3.  You have to remember that the optimisation process is purely a mathematical process which can't detect geometric accuracy.   Depending on your images, you may be actually introducing a small amount of distortion to make the alignment of points closer together overall.

[Marcel]

We've had 'staircasing' artifacts as well (we nicked it "stepping"), but it was usually just small that were raised a bit higher than the rest of the scan.

Our interpretation was that it had to do with an incorrect alignment of the cameras as well (ie. a camera pair that had an incorrect position but still contributed to the Dense Cloud).

Now we manually go through the cameras after alignment, and remove any cameras that are too close to each other. We also cull the sparse cloud and then run Optimize Cameras. That (and better quality images) seemed to have gotten rid of it.

[Drakhain]

Thank you, i'll try to make less overlap between my shots and see the difference.