Reality Capture fails doing large terrain models

OK- this is NOT a RC vs PS post.

I have done some great modelling with RC recently- it is amazing software and I estimate about 5x faster at processing than anything else I have used. However, while it works really well on smaller areas ( up to couple of hundred metres long and wide) when I try and run larger areas it inevitably fails to align properly and I end up with warped models and large errors in Z. I am using ground control points (surveyed using RTK GPS as minimum, or total station/ digital level- so nothing wrong with my control).

Perfect example from earlier today. UAV mapped area of ~1700m x ~300m wide. 6 GCPs spread evenly over area. I locate GCPs in images (set image GPS exif data to "unknown"to disable) and RC aligns images pretty quickly but it gives me a warped surface and big Z error at one end- in 2 GCPs:

Tried a couple of times, but same result.

Same images, similar process with Photoscan (although the GCP location is far superior to RC… have I mentioned this before?? :slight_smile: )
Gets me the real result:

Am I doing something wrong in Reality capture?
Is there a setting that would fix this issue?

I really want to be able to use RC to do all my modelling and orthophoto production as it is so fast and from what I’m seeing I get more accurate terrain models (in smaller areas anyway). Just hasn’t worked for me when I try to run long skinny projects (and since I’m a surveyor doing roads and railways etc- this is what I am working with most of the time).

Anyone got any advice???

Hi peterwatson

  1. What is your global OVERLAP on images ?
  2. what settings are you using for ALIGNMENT ? please post a screenshot of just the alignment settings
  3. What reason do you have for DISABLING the GPS information in the images ?

Hey,
Which UAV/Camera did you use and what are your alignment Settings in RC.
I saw a similar problem and I’m trying to find the cause.

Milos,

  1. Do you mean overlap set for image capture? or is this a setting in RC? From memory I was using 70% frontlap and sidelap for image capture.

  2. Alignment settings:
    Image

  3. Disabled image exif GPS data as per this post:
    Turn off location data from images?

As in other software assume it is necessary to turn off the approximate GPS (not RTK) locations in images when using GCPs with far more accurate locations? If I do alignment with image prior pose set to “location” and also have GCP locations I get this sort of result:

Not sure if this just due to inherent innacuracy of camera GPS position vs GCP accuracy? or different coordinate systems -WGS84 lat/long for images, and UTM easting/northings for GCPs ??

Gal Davidson- using DJI phantom 3 pro, alignment settings as above. Have had this phenomena occur many times- but it’s pretty much always on long skinny corridors eg. roads.

Hi Peter Watson,
I use Phantom 4 and I don’t have this problem at all.
I have had this problem when I tried using my mobile device with much fewer shots - about 140…
I’m trying to improve the settings to get better Alignment and without the “Distortion” Bug.

Currently, I’m playing with the “Max feature reprojection error” - Trying to extract smaller details as streetlight that disappears.

Capture.JPG

Thanks, I’ll give this a try. Have you tried processing longer runs… eg >1km.
I seem to get good alignment with smaller areas, but when I try and align bigger areas it fails to align correctly- using same altitude/overlap etc.

eg.

https://youtu.be/_6lcKuUlCE4

I found that some photos cause problems, so what i tried to do is to align 20% of the Object/model let say the first 200m of your road and after alignment is good I add more photos and do align again until I found where is the problem. (probably some photos with an angle or lack of overlap)

I’m coming from the 3D scanning industry and what could be a nice feature in RC it’s if after alignment there could be separate groups of point cloud that had a problem of alignment and we could decide if we need them or not and then register them.
Currently that way of work that you will receive the biggest model or a broken(with distortion) model that consist of different parts that not aligned well - it very frustrating to start all over again - the best to put it in separate models and then with cloudcompare i could merge them.

What I mean is: many times you have some photos that couldn’t align especially if you combine pictures from the Drone and from the SLR (From the ground) - what happening in the background that there are a few Segments (Groups of Tiles Points) and the software doesn’t really know how to align between 2 different sets of photos. (I’m talking about the automatic process with GCP)
The Solution in the 3D scanning industry is doing it Manually or Semi-Automatically - register that Group of clouds and receive the full model.

I can explain better if the developers would like to improve that alignment of different sets…

Hi Gal
In short :smiley:

The main issues can be pinpointed to not enough overlap and sharp angles. From the screenshots that I have seen I can say for sure the overlap is not the best. try to get to 80+% overlap to get better bundle stability ( alignment of the scene ) and I can say from my experience even misalignment gets away…

regarding ALIGNMENT settings, set these settings, then delete all COMPONENTS already in the scene and realign it again.
if it is a small sensor, I would recommend you to change the DISTORTION model from BROWN3 to Brown3-4 with Tangential ( any with TT distortion from the menu )

Alignment_settings.jpg

Milos,

Changed settings as per your post but still same distorted output:

Might try flying area again with 80/80 overlap

I have had pretty good luck shooting orbits rather than grids.

I’d suggest keeping your photos from the grid pattern and add and few over lapping orbits.

I’ve done 12km x 12km like this (from heli, not drone though)

Thanks Chris, I might give this a go first. Also there is another app I’ve used called Altizure which allows angled gimbal shots instead of straight down/nadir photos- good for getting detail of tree trunks, building sides etc. I’ll go and do some more flying with both methods and see what I get.

Hi,

I have also had problems with misalignment in RC on data sets with a large amount of overlap. Data sets that have been processed through PS without any problems. However significantly slower.

Is this a common problem?

Sorry I know this isn’t a RC vs PS post; however I would really appreciate any help with this!

Cheers Jarrad

altizure (dose that work with phantom 3?) might be a go.

i think angled gimbal shots will work better than straight down.

you will also want to change heights. don’t do everything at the same altitude.

I haven’t really had time to do too much experimenting with shooting patterns with uav. but I’m interested in solutions for setting up flights paths, so feel free to let me know once you have found a good one.

jarrad,

for the alignment to work well you need to see the same object in the start and end of a loop of photos. this seems to work best if they are smaller areas, and more of them. rather than big long sequence before you see the start again.

also you will need to experiment with some alignment settings like, image overlap, max feature per image, detector sensitivity.

Had the same issue but with an elevation of a 75 m long straight wall.
Wishgranters tip to delete all components and calculate again worked perfectly!
The worst deviation to my theodolite-measured GCPs is now 5 mm and that is quite good since it’s in a difficult spot…