Sparse cloud is clean, but dense cloud kind of a mess? screenshot attached.

[joacampb]

Hello,

At the dense cloud phase I am getting wild results I've never seen before. the sparse cloud is clean after optimization and gradual selection, but the subject matter in the dense cloud is surrounded by a chaotic mess of points. Please see attached images for visual which also includes the chunk info and analysis showing all the processing options I used.

I did not use Masks when processing, which I assume would resolve some of the issue. but I am curious to know why this happens at all.

About the data set:

• The source images are from a Canon 5D MK IV with 35mm, 50mm and 100mm lenses ( testing 100mm lens for fine details on object ).  All camera and lens combinations were imaged from roughly the same measured distance, but this was not a turntable setup. so the camera moved around the object and there is some minor variance in the distance between subject and sensor among the images.
• The RAW files were converted as 16bit tiff using Rawtherapee and AMAZE demosaicing. Only adjustments were to lighten shadows and minor input sharpening to remove blur from anti-aliasing filter built into camera.
• The sparse cloud is cleaned via Gradual selection. First reprojection uncertainty ( threshold 10), Second projection accuracy (threshold 3.0), finally reprojection error is iterated until Max RMS error for the chunk is 0.3 or less. Between each gradual selection filter step the cameras are optimized.

When I run the dense cloud with Aggressive filtering, it helps, but doesn't fix completely. I am used to getting some 'fuzz'  in the dense cloud but never like this before.

If I can do a better job of explaining the issue then please let me know. Attached screen shots are here to help.

Thanks!

[sfmdd]

Hi, i'm not sure if this helps with your problem:

1. Looks like if you're using way to much images for this kind of an object, around 50 would do the job. Maybe the software is also messing around with the different focal lenghts in close ranges but i'm not sure about that.
2. Don't adjust images at all. I'm also a bit confused about the anti aliasing filter in the RAW images, there shouldn't be any filter. If so: Canon, what are you doing?

[Dieter]

Since the latest versions, I am no longer satisfied with the dense point cloud. The basic setting of the images to be evaluated in parallel is too low.

You can change this by making an entry under the advanced settings: "BuildDenseCloud / max_neighbors = 1000"

I work with 1000 to get good results. Since you only have 864 images, you can play around with the settings here. A "-1" always means evaluating all images for the dense point cloud in parallel. Then you have the best point cloud, but the evaluation also takes the longest.

I would also set the filtering to aggressive.

[Alexey Pasumansky]

Hello joacampb,

I can suggest to try the version 1.7.1 and report back, if you are getting any better results. Using the tweak, that Dieter has specified above would also help, if you have excessive image overlap (most likely, setting it to 200-300 would already be sufficient).

As an alternative approach you can generate mesh model in 1.7 using depth maps based reconstruction option and then call Tools Menu -> Dense Cloud -> Sample Points command and sample the dense cloud from the generated mesh, if it doesn't show any unwanted artifacts.

[joacampb]

Thank you Alexey and thank you Dieter!

I updated to 1.7.1 and applied the advanced setting tweak and it looks great. No more fuzz.

follow up questions :

Is there any general rule of thumb for setting the max_neighbors setting?

and when using python, would there be any negative side effect (other than processing time) of setting this value to something like 1/3rd the total image count? So if I have 900 images i would set max_neighbors to 300 programmatically for example.


Thank you again!

[dronesurveyconsultancy]

Hello,

I have the sampe problem, how do i put in the tweak for the max neighbors, because its not working and giving errors.

Br