Dense Point Cloud Gaps

[Jholland]

Hi Everyone,

I'm trying to create a dense point cloud using a sparse cloud that has been aligned and thoroughly cleaned up.  All of the cameras have remained calibrated and all images have a high quality score so I was expecting every photo to contribute to the dense point cloud. But I’m getting large blocky gaps in the dense point cloud that seem to reflect photos and not features on the ground.   

The images were captured with a drone using a grid pattern.  I’ve processed the dense point cloud using ultra high, medium and low quality settings (mild or no filtering).  The ground surface is a variety of forest types which have produced points elsewhere, so I'm baffled why these areas are missed.  Any help would really be appreciated. 

Thanks,
James

[Alexey Pasumansky]

Hello James,

How does the camera alignment look like for the given area? Can you provide a screenshot of the tie point cloud view with the enabled camera locations?

Also please specify the matching/alignment parameters used.

[Jholland]

Hi Alexey,
I've attached images with the alignment parameters resulting point cloud after edits, the coverage of the aligned photos, and a close up of the thumbnails which seem to coincide with the gaps. 
Thanks
James   

[RHenriques]

Hi Alexey

I'm experiencing the same problem as James in this latest 1.8.4 beta.
The survey was carried out with crossed stripes (crossed grid pattern) so overlap is not a problem.
However, sometimes when there is people or things moving around, some masking helps a lot. These latest Metashape versions seem to be too sensitive to random object moving. In a survey I've masked 4 persons that were moving around and things improved a lot.
Best regards

[Alexey Pasumansky]

Hello Renato,

According to the screenshot, James is using version 1.6.5, so the problem might be different to the one you describe. But if possible, please share minimal dataset that can be used to reproduce the problem (and the project with the masks applied) to support@agisoft.com.


Hello James,

Have you applied any strong gradual selection filtering to the tie points before building the dense cloud? According to the screenshot only about 5% of tie points are used (marked as valid) in the chunk.
You can try to "return" the points by executing the following single line command in the Console pane and then re-building the depth maps and dense cloud:

Code: [Select]

Metashape.app.document.chunk.triangulatePoints()

[RHenriques]

Hi Alexey

I've been testing this issue a lot and it seems to be linked to objects moving around (a few persons).
Today I've masked them all and things did improve a bit. These latest versions seem to be too sensitive to moving objects.
I've also increased the Rolling Shutter correction to the most aggressive one, so I cannot filter which was the biggest culprit. Probably the two.
I'll send you the dataset that I'm using by email via a PCloud link.
Best regards

[JamesHolland]

Hi Alexey

I have definitely refined the sparse points a lot, trying to follow the steps and to get the recommended accuracies of a USGS methodology  (https://pubs.usgs.gov/of/2021/1039/ofr20211039.pdf).  The final accuracy in the Reference View is .009m, and .19pixels for the Chunk which is close to the recommendation.   I've attached an image with the processing steps and parameters.  For example, I filtered by Reconstruction Uncertainty 3 times, using 110 as the first threshold.  Projection Accuracy was used twice, and Reprojection Error was used three times.  The Chunks that were not used are stroked out in red. 

I had assumed that if the cameras were still calibrated at the end of the sparse point filtering, they would
be used for the dense cloud generation.  I can go back to one of the previous chunks during the sparse could filtering to generate the dense cloud, but the accuracies are not nearly as good until I get to the final filtering step and camera realignment.

There will be some movement in these images at the tree tops, but I was hoping the final sparse point cloud would be solid ground points. 

Thanks in advance for any suggestions.  I will look into upgrading the software. 
James