Shouldn't the dense pt cloud exactly match GCP's at the GCP locations?

[jmaeding]

I've been using PS Pro for about a year now, and have generally gotten really good results using both "targeted" GCP's, and points I "photo ID'd" after the flight and got coords from previous survey. I fly a DJI P4 Pro V2, very typical setup.
Today, I flew a job and generated GCP's as normal, such that they are 2x flight height apart. I flew at 250', and have GCP's 500' apart.
My workflow is that of the USGS, as written about in these forums.
I align photos, convert to CA83-VIF, eliminate low quality pairs...all that, and then uncheck the images.
I bring in control coords, tell the photos where the GCP's are, and eventually build dense point cloud.
My question is:
Once I get the dense point cloud, should the points right at the GCP's exctly match the elevations of the GCP's.
I thought in the past they had. Today I got points as much as 1.2 feet off. Then some dead on.
I know PS involves a lot of factors, but assume I have very good imagery, good overlap, and so on.
I'm trying to figure out proper PS behavior so I can troubleshoot problems with me, or it.
thanks

[jmaeding]

welp, got no replies, but I think the answer is no. The point cloud will not match the GCP's exactly, there is some kind of fit going on.
As I was investigating this on a particular job, I read how you can set mode to "estimated" and see what the software got for each control point after processing.
Then I realized I should be paying attention to the error values before processing, as those will indicate the difference between the GCP numbers and where the model thinks they should be (after registering markers on the photos).
In the end, the lesson I learned is if you get error values higher than 1 ft, there is a mistake in the GCP coordinates, or the photo registration.
One typical mistake is duplicate elevation values in the GCP's from copying and pasting numbers wrong.
Another is placing markers on a spot that is not the same for a given GCP.
I added a step (19) to my essentially USGS workflow, which is summarized here:
Condensed Workflow

1)   Make project and load photos, leave checked
2)   Convert to State plane, ESPG::2230 = Zone 6, ESPG::2229 = Zone 5
3)   Check that GPUs are set to be used in Tools->Prefs->GPU tab
4)   Align Photos, accuracy on high
5)   Show projection and error columns in photo list
6)   Do optimize (not all boxes checked)
7)   Do Gradual Selection using Reconstruction Uncertainty, select 20% at a time, goal is 10. Repeat 6 and 7 until you get close.
8)   Do optimize (not all boxes checked)
9)   Do Gradual Selection using Projection Accuracy, select 20% at a time, goal is 2.5. Repeat 6 and 7 until you get close.
10)   Open reference settings and set Tie Point Accuracy to 0.1, Do Optimize (not all boxes checked)
11)   Uncheck photos as we are done with using their gps tag data
12)   Create Control Points list as csv
13)   Import csv to create control point markers, say Yes to All when asked
14)   Right click on a point by marker, and choose Filter photos by point.
15)   Move marker to correct location in 2 photos. This refines the locations.
16)   Repeat right click on a marker, and choose Filter photos by marker this time.
17)   Go through each photo, and if target is clear in the photo, drag the marker to the correct location. Grey flag means “not used”. Green flag means “used”. Hit page down to go to next photo. It should be click, page down, click, page down…
18)   Block any markers, using right click, that you decide are not clear, if already set green.
19)   Double check that none of the markers have high error values. Also look for duplicate elevations indicating the elev was copied from another spot on accident.
20)   Do optimize (all boxes checked)
21)   Do Gradual Selection using Reprojection Error, select 10% at a time, goal is 0.4. Repeat 6 and 7 until you get close.
22)   Build dense point cloud medium level
23)   Build DEM (needed for ortho)
24)   Build Ortho photo (machine)
25)   Export Laz point cloud file
26)   Export Ortho photo as jpg with world file. Set compression to jpg and select jpg in file type in dialog.
27)   Create .rcp file in Recap for Autocad/Infraworks use.

[Alexey Pasumansky]

Hello jmaeding,

The 3D location of the GCP (marker) in Metashape Pro window is estimated according to the marker projections on the source images. It means that the 3D point is defined by the exterior and interior orientation of the cameras and could be estimated even without the dense cloud.

The simple example, when GCP location may be considerably different from the reconstructed dense cloud / DEM: if the point is measured on top of a very thin and high pole, which cannot be reconstructed due to the selected depth maps quality parameter.

[MaciekK]

Hi everyoneI'm hooking the thread because I have a similar problem. The dense cloud does not reflect the correct heights at the control points. I get very good alignment results. Total error at the control points about 1 cm, at the check points about 1.5 cm. After generating a dense cloud, I get heights at control and checkpoints different by about 4-5 cm (all values with the same + sign) from those measured in the field with a total station. Job made by Phantom RTK, 65m AGL, gsd below 2 cm. All camera positions are fix. After alignment, the total error of cameras about 3 cm. After generating a dense cloud with the ultra hight parameter, the situation improves slightly, errors fall to about 3 cm. But that doesn't satisfy me anyway. Is this the extreme accuracy of MS dense cloud generation? A bit strange because aerotriangulation comes out about 4 times better. Do I make a mistake?