Strange Noise in Point Cloud on new Desktop PC

[4DS Niklas]

Hello everyone!

For my latest Project, i need to document the facades of a church, specifically one of its 50m high towers. For that took around 900 pictures from different perspectives all around the tower.

I used control points in the form of markers which were already on the facade. I measured those in a local system using a Leica TS16 Totalstation. I got pretty good results for the transformation via those points (around 3mm error using about 10 points).

Now to my problem:
We recently transfered our metashape license to our new computer. It has a AMD Ryzen 9 5950X 16-Core Processor 3.40 GHz CPU, 128 GB Ram, and a Nvidia RTX 3090 GPU.
Now, on our old machine, we got a pretty good result. We processed the project again on our new pc, using only slightly different control points to change the coordinate system, but now, we have some weird noise around the edges of the tower which i can not really explain. Anyone got any idea what the problem might be?
Please tell me if you need any more Info! See the picture attached to see what i mean.


Thanks in advance!

Niklas

[Alexey Pasumansky]

Hello Niklas,

If there is an excessive image overlap, it may be required to raise the number of neighbors to be considered for the dense cloud filtering (can be done only via Tweaks).

However, I could suggest to try building the mesh model using the depth maps source option in Metashape 1.7 and check, if you get good quality model without any similar noise.

[Brandon]

I have been experiencing something similar but somewhat worse, I have a new machine, Ryzen 5800 Nvidia 3090 128gb of ram, and have been reprocessing datasets. Formerly clean results have a lot of garbage that looks like it comes from the background.

[Brandon]

I would like to ad that this garbage is not visible in the sparse point cloud.

[Alexey Pasumansky]

Hello Brandon,

Can you share the source image set for the problematic project to support@agisoft.com together with the project file, containing the alignment results (cameras and tie points) with the depth maps?

Additionally please describe the image acquisition scenario: whether all the images were taken simultaneously by the synchronized camera rig, or if the single camera approach has been used? In the latter case, whether the person has been standing still and camera was moving around, or if the person has been rotated in front of static camera?

[Brandon]

There is a low limit on the size of the images I can post here how are data sets usually shared?

The example I posted was taken with a single camera, but I have another project that was taken with an array that exhibits a similar result. Dense point cloud calculated with earlier versions are clean, Point clouds calculated with the current version have this "fringe"

[Brandon]

I shard the footage to support@agisoft.com using dropbox

[Alexey Pasumansky]

Hello Brandon,

In the given case you can clean up the dense cloud using Confidence filtering (select points with 0 -1 or 0 - 2 filter and remove the points, then reset the filter).

However, if you are interested in mesh model as a final product, you can build mesh using the depth maps source (skipping the dense cloud generation) and it will not have any similar noise around the actual model.