ContextCapture not respecting GPS locations (bad Aerotriangulation)

Really strange indeed. Which Contextcapture version you are using?

But you are using AT settings wrong. You don't set both Final rigid and adjustment constraints to photo positions you have to choose one. Final rigid is for innacurate.

Under Main settings you can lock positions or use "Extend" option for rotations. It is advised to allow at least rotations to change as lot of photos will be drop if they both are locked together. Regarding verification of locations each software has slightly different results so there really is no ground truth. If this is Nadir flight then reusing calibration from 3D project could also improve accuracy.

So after some digging, I think I have a better understanding of what's going on:

• Only 1 of (Adjustment Constraints or FInal Rigid Registration) should be selected. Never both. No idea why you are allowed to select both.
• Adjustment Constraints should only be used for high precision coordinates (<5 cm). So we should just use FInal Rigid and leave everything at default.
• The only Poses and Tie Points option available during initial AT is 'Compute'
• The other options (Adjust, Extend, Lock) become available on subsequent AT runs and vary depending on whether or not all images were included in the Main Component.
• Using 'All images' on subsequent AT runs allows you to 'Extend' the current Main Component by attempting to re-add those photos that didn't get included initially.
• You can also run 'Adjust' to help fine tune Position and Rotation of existing, good images
• Keypoint density should be set to High for lower resolution photos (which we have)

I don't want to get my hopes up, but I got good results for the first time. *fingers crossed*

Oto Appreciate your assistance with this.

So after some digging, I think I have a better understanding of what's going on:

• Only 1 of (Adjustment Constraints or FInal Rigid Registration) should be selected. Never both. No idea why you are allowed to select both.
• Adjustment Constraints should only be used for high precision coordinates (<5 cm). So we should just use FInal Rigid and leave everything at default.
• The only Poses and Tie Points option available during initial AT is 'Compute'
• The other options (Adjust, Extend, Lock) become available on subsequent AT runs and vary depending on whether or not all images were included in the Main Component.
• Using 'All images' on subsequent AT runs allows you to 'Extend' the current Main Component by attempting to re-add those photos that didn't get included initially.
• You can also run 'Adjust' to help fine tune Position and Rotation of existing, good images
• Keypoint density should be set to High for lower resolution photos (which we have)

I don't want to get my hopes up, but I got good results for the first time. *fingers crossed*

Oto Appreciate your assistance with this.

Also you may try to run with photo downsampling applied. It lowers resolution for AT but may work faster and produce better results and more tie points when using high resolution cameras. After AT downsampling can be removed.

Mike Wilmington said:

• Keypoint density should be set to High for lower resolution photos (which we have)

Have noticed the opposite that high density produces less tie points but drops more innacurate ones. But helps when flying over woods. Check automatic tie point distribution on photos after AT to get idea which points it finds.

Mike Wilmington said:

Adjustment Constraints should only be used for high precision coordinates (<5 cm). So we should just use FInal Rigid and leave everything at default.

Yes but you said you have accurate positions. So you set this but under adjustment settings allow position to calculate. 

Latest version has some improvement related to this. Confusing I know but gives you more control over whole process: communities.bentley.com/.../update-20-1---revised-at-engine

We have standard GPS positions. As far as I know, that's not going to give me <5cm precision. I'm actually not even noticing a difference when selecting "adjustment constraints" and "final rigid registration" in most some cases. Below is a set of captures going back and forth, capturing a scene from many angles to ensure there is a lot of overlap. I tried exhaustive pair selection and adjustment constraint in this case. You can see it totally trashes the locations of the points. This is a collection from ground level with a standard GoPro (4K images). If I just do a single, continuous video and extract that image set, it will work most of the time. Combining multiple sets even though they have similar overlaps in images almost never works.

I just don't understand why the positions would be allowed to move 20+ meters.

For my testing, I'm trying to capture a street. I go up and down the street and try to capture enough angles that the overlap should be obvious (N, NE, E, SE, S, SW W, NW). But I must be doing something wrong. Maybe my overlaps are not good enough. Maybe my fov is wrong. I know I have enough images because I'm using 0.5 meter spacing so there are a ton of images.

Maybe it doesn't like linear paths? The data is usually not oriented correctly (z up) when I do linear paths... even though it's specified as WGS84 and the height is present in the metadata.

I'm at a loss...

Ground truth (GPS)

Final AT

AT final position vs metadata

Settings for the run above:

A. Positioning/georeferencing:
- Adjustment constraints: photo positioning metadata
- Final rigid registration: none

B. Main settings:
- Key point density: Normal
- Tags or QR Codes extraction: Disabled
- Pair selection mode: Exhaustive
- Color correction: blockwise
- Splats: Enabled

C. Estimation policies:
- Tie points: Compute
- Position: Compute
- Rotation: Compute
- Pre-calibration stage: Enabled
- Focal length: Adjust
- Principal point: Adjust
- Radial distortion: Adjust
- Tangential distortion: Adjust
- Aspect ratio: Keep
- Skew: Keep

Try "legacy engine" under low-level AT settings.

But your explanation is confusing as at first you said GPS positions are accurate and verified and not just standard GPS.
Are all these images taken with same camera?

I just saw this “Maximum position metadata confidence.cfg" from your post and I'm trying it now.

Sorry if I implied they were perfect or near perfect. I didn't expect them to move so much. They are standard GPS coordinates and are of reasonable enough accuracy from what I can tell.

I will also try the Legacy Engine option in addition to "maximum position metadata confidence.cfg." I just need to find out where that is.