Issue: JPG alignment, arched ground, bad elevations, distortion curvature

[rttgnck]

I have been plying my Mavic Pro and collecting plenty of test projects. However, on a few of my projects (larger sites) I have been working on I noticed that after processing and rectifying my clouds and comparing to elevation information gathered conventionally that my clouds (and camera alignment) have an arching/curvature distortion in one direction.

My control point rectification was done manually with CC, and the areas nearest them is close to expected, it is when I start to get towards the edge of my cloud that things get worse and worse.

After checking a few known locations elevations they are close, but about 1/4 from the edge of the cloud I have the curvature distortion resulting in an elevation error of 8ft.

On a previous project that showed me this issue my control points were good, 4 corners surrounding a site, but the center of the sight was as high as 5ft above expected in some places.

At this point I am not sure where this distortion is coming from and the only thing I can thing to try next is

1. use DNG format instead, maybe the GPS elevation in JPG is just to bad (to bad DNG writing would slow the job down from what I have read.)
2. calibrate my camera for photoscan using the agisoft lens application

If anyone has run into this issue please help. At this point the information gathered is essentially useless because I have a curvature in my project that does not exist in reality.

Rest assured that this is not normal earth curvature as the site is relatively small (10 acres) and I have lots of ground elevations to compare against, at both sites.

I have attached an image showing a side profile to illustrate the issue.

[SAV]

Hi rttgnck,

The curvature is probably related to a) processing settings, b) image overlap and or c) image quality.

Regarding a):
Choose MEDIUM or HIGH for the Accuracy and REFERENCE for pair preselection. From my experience I have seen that lower accuracy settings are increasing the 'arching' problem.
In the reference settings, change the camera accuracy from 10m (default) to 3 or 4m. From my experience the GPS tags of pictures from DJI drones are always more accurate than 10m. Mostly they are around 2-3m. By changing this value you will help to reduce your total error.

Also don't forget to do run the OPTIMIZE CAMERA ALIGNMENT (magic wand button) after you have cleaned up your sparse point cloud, e.g., manually removing 'outliers' and inaccurate points. EDIT > GRADUAL SELECTION will help you to achieve best results.


Regarding b):
Something between 70-90% of side and forward overlap (depending on scene complexity) are necessary to achieve good photogrammetric modelling results. If you have a lower overlap than that, you won't be able to get a geometrically well constrained reconstruction.

Regarding c):
Blurry images are BAD. High ISO/noisy images are BAD too. Choose the right camera settings (to reduce noise and get well exposed images) and horizontal flight speed (to reduce motion blur).

Note that your elevation from your model might still differ to what you might find in maps or online. This is due to the fact that drones from DJI store the 'altitude relative to the point of take-off' in the images. For example, if you took off on a hill that is 500m above sea level, that would be 0 elevation in the geotags of your images. There are, however, ways to fix the DJI altitude/elevation problem. Search the forum 

Last but not least, if you still have warping/curvature problems or issues with the elevation, use accurate ground control points (e.g., surveyed with an RTK GPS or total station).

All the best.

Regards
SAV

[karad]

I think your model exibits the well known "dome effect" (see literature), caused by insufficient calibration of the camera..

[rttgnck]

Thank you for the replies.

Hi rttgnck,

The curvature is probably related to a) processing settings, b) image overlap and or c) image quality.

Regards
SAV

As for your concerns
a. I will look into this to see if it helps
b. I was using 90/60 overlap (default in DJI GS Pro), thought that might be it and attempted to recapture the site again (before your comment) with 80/80 and got the same results.
c. I am still learning my camera settings, but from what I can tell they are good images, could probably still be better.
Note. I have adjusted my point cloud using ground control points with a good XYZ (survey monuments) in CloudCompare. This fixed the elevations being focused around 0 and brought them to the surface, but the "dome effect" caused areas that weren't at or near the control points to be off from the expected elevations.

I think your model exibits the well known "dome effect" (see literature), caused by insufficient calibration of the camera..

I will look into the literature, I did a quick calibration yesterday after my post and still got the same results, but maybe I just had bad calibration.

What I find weird about this is that DroneDeploy has given the same domed results (albeit I am on trial or free) which is why I thought it was in the data acquisition.

[SAV]

Hi rttgnck,

If you don't mind then please share your data and I'll have a look at the issue.

Without proper ground control points one would expect a little bit of 'doming', but in your case it looks quite extreme.

Regards
SAV

[rttgnck]

Hi rttgnck,

If you don't mind then please share your data and I'll have a look at the issue.

Without proper ground control points one would expect a little bit of 'doming', but in your case it looks quite extreme.

Regards
SAV

Is it possible that the camera itself needs some sort of calibration? Seems unlikely. Do GCPs help the initial alignment and calculation to calculate at "the surface". If I don't incorporate GCP before the initial alignment it's almost like it does not know where it is and just calculates based on arbitrary units, scale, and elevation. Like my scale is many times smaller than reality, and the elevations tend to be +-30 of 0.

Even if I did share data, this has happened in all of my data sets so far (4) that are larger than just my house. Some doming occurs in all 4 sets when processed without GCP and then rectified after. Of course the doming occurs immediately and is not a result of the rectification. Further none of my data sets have visible GCP markers, I am hand picking them based on their expected location and a cad drawing of the site.

[Jeremiah_ROWE]

Please upload a screenshot showing the camera positions from the top looking down.

[rttgnck]

Attached is a top down view. This is with 80/80 overlap (my most recent collection). Unfortunately it looks like some photos had a serious drift when in the air. I noticed it was a little windy that day with medium satellites and saw the quad drift a little on a few photos. I do not believe the drift is at fault here as I have 3 other data sets with perfect path alignment and they also have the doming.

[Phillip Marx]

Has this issue been resolved?

[lyhour]

I meet this problem as well. Does this problem solve yet ?

[Eric Baird]

Are you by any chance using a fisheye lens with the Mavic?

If so, you might need to go into the Menu/Tools/Camera Calibration... page, select the block of photos taken with the fisheye, and manually change their "Camera type" from "Frame" to "Fisheye" using the dropdown box near the top of the screen.

Metashape seems to be great with fisheye lenses, but you have to remember to manually select the option.