Collectively curved images and consequently curved models.

I have made many terrain models, with good results, using a Phantom 3.  I was testing some new parameters for image collection that I wanted to try, and the result produced an extremely curved model.  The curve in the model, however, was caused by a curve in the photos.  I flew a grid that had 15 or so back and forth lines, and I had enough overlap so that 1 automatic tie point was “seen” by 4 lines of photos.  The height of the phantom was consistent throughout the flight.  The software placed all the photos in the correct location, but the photos as a collective whole looked as if they were pressed over a ball: the center was high and each successive photo sloped down towards the edges.  The curvature was so extreme that I immediately noticed it after the initial processing was complete.  The curve is there with or without using the geolocation information.  I haven’t tried GCP’s, but I’m more interested in why the curve is there.  I went back and look at my earlier “successful” models, and noticed the same curve in almost all of them, only to a lesser degree.   

Jamie

Dear Jamie,

We have noticed that the new DJI cameras (X3 and X5) are facing two challenges:

A. The computed focal length could get two solution mathematically; however one of them does not make sense physically. In order to select the correct computed focal length, we included another option for the calibration. To select this option: |

  1. Click Process>Options.
  2. In 1. Initial Processing, make sure that Advanced is selected.
  3. In calibration, in internal camera parameters, select All Prior.
  4. Process step 1.

B. Also, we have noticed that the images captured with DJI cameras suffer from the rolling shutter effect that introduce artifacts and errors in the processing. We have modeled the rolling shutter effect and the model is implemented in version 2.1. For the moment, you can fly slower in order to reduce the rolling shutter effect.

Best regards,

 

This artifact is also present with Phantom 2 Vision+ with and without control.

 

Reprocessing with 104 release and your suggestions.

Thanks for the help.

 

Some of my results are very similar to the image posted above.  The project that initiated the post had much worse curvature.  I too am reprocessing with the suggested settings.  

Back to the drawing board.  Notice the blue dots (initial computed position) are on the same plane.  It’s the green dot (optimized positions) that are being warped.  My flight was 150 feet high and I had between 80% and 85% side overlap.

Looks like you need GCP’s.

If you don’t have access to Survey Grade GPS,  there are some recent cheap DIY solutions for RTK GPS.  

L1 RTK GPS boards are only $50 + $50 antenna.  It takes a lot longer, but cm accuracy is obtainable.    

Dual Frequency “Survey Grade” GPS Boards are only $800 but obviously require more expensive antennas.  

The cost for centimeter RTK GPS is dropping if you don’t mind the DIY aspect.  

I’ve been looking into RTK solutions for couple years now, and yes some open source projects have made progress.

Piksi is one of the projects I’ve banged my head against for a while now.

So far the most versatile system for me is Bad Elf because the units are very small and are put on aircraft and ground targets. Most open source systems either lack sufficient hardware and/or software for efficient surveying and processing of signal data (Rinex protocal).

Using professional survey systems are by far the best way for GCP acquisition albeit a 20k-50k dollar addition to your kit.

Perhaps we should open a thread here on this subject. I am interested in how others are acquiring high fidelity GCPs and their work flows for incorporating into programmatic outputs.

Back on subject: After using suggested processing configs I still see DSM distortion. I am sure I’ve induced this error somehow!?!?

Yes, the suggested solution didn’t help me either.  I did produce a successful model recently, but I don’t have access to that computer at the moment.  I will post the parameters of that model tonight.

Going off the GPS suggestions, has anyone used the Reach GPS system.  It looks intriguing, but I do not have RTK experience: so I don’t know how viable of a solution it is.  

Yes looked into it…

The Reach receiver is limited in performance for the XYZ positional fidelity we need at the GSD we sample at. I would like this system if better logging implementation.

uBlox receivers are widely used and are good bang for the buck but I still don’t see a efficient implementations for use outside pro-gear.

I started a new thread to discuss RTK GPS.  I apologize for hijacking this one. 

https://community.pix4d.com/t/4703-Discussion-on-RTK-GPS-Systems- 

Hope to see you there.

I seem to have made some progress with the distortion problem, at least with the phantom 3 camera which, I think, is the same as the X3 camera.  

Go into Image Properties Editor>Edit>Edit>Estimate from EXIF> Then change Camera Model with Distortions to 0.

I am processing the point cloud now, but the photos and Tie Points seem to be undistorted.

Interesting.

This is my fault for not characterising each of my sensor systems carefully and assuming that the lens correction tables in Pix4D are accurate enough. Pix4D is a well rounded work-flow tool in photogrammetric processes, but do diligences must still be take in every aspect of the data acquisition to deliverables. Or said another way, (I need to be less lazy)…!

 

Hello. I’m hoping that someone here can talk about this subject again. 

  1. Jamie: I tried your solution to setting radial distortions to 0 for the Phantom 3’s camera model, and I indeed produced a better model. However, I don’t know if this is correct. Surely there are still some radial distortions present in the images unless the camera’s JPG output has been undistorted already. Any progress in understanding this issue?

  2. Garry: Do you have a method for characterizing a sensor system in-house? I’m curious about it because I’ve been gearing up to create my own camera model with OpenCV using camera calibration tools.

  3. Pix4D: I put in a support request asking about this issue. The reply I got was not any different than the one given here in this forum on Jan 16th. Because I also implemented the suggested solution, but I still appear to have a curved model, I’m not sure the “correction tables” in Pix4D are correct. 

We need to know a solution to this problem, or the Phantom Profession 3 camera might not be considered a good mapping tool using Pix4D.  Please advise. 

Dr. Stephen Scheidt

Stephen, I’m not sure the solution is correct either, but I haven’t tried playing with the radial distortions any further.  One thing that can be tried is to take oblique images with the orthogonal imagery; I have read a couple of articles that found that helped with terrestrial photos.  I think the only way to be certain that a model is not warped is to use GCP’s shot with a laser transit or RTK GPS.

Jamie 

Hi all,

Sorry for the late reply.

Pix4Dmapper 2.1 was released in March, and it is now possible to apply a correction algorithm for the rolling shutter of some cameras like that of the Phantom 3 drones. Note that this algorithm is mainly effective for grid projects.
Please see our step by step procedure here: https://support.pix4d.com/hc/en-us/articles/208460436

If this does not work, you could apply the All Prior settings as described by Despina above (option A), alone or together with the rolling shutter correction.

As highlighted by Jamie, inserting GCPs in the model will contribute a lot in solving the problem because it will be adjusted to the GCPs and forced to be flat. That would be ideal.

Regards,

Hi! I’m interested in this post. Having a Phantom Pro 4 I have notice the same issue observed by Jim. Could be possible that flight speed is affecting the results as it was pointed in a previous reply? thanks in advance!

@Mauro

In some cases, especially when flying low and fast with the DJI Phantom 4 Pro camera, the curving of the scene can appear. 

I would recommend reading the following community post in order to get additional information and potential workarounds: Phantom 4 Pro Rolling shutter effect.

1 Like

The structure-from-motion doming/dishing [elevation] error is well known and well documented - see M R James research which calls for gently curved, convergent, non-traditional (non-linear/non-parallel) flight lines along with improvements to the camera’s calibration and some oblique photos at “different” heights.
I have been flying curved flight missions for the past couple of years and noted that the curved flight lines provide a) more diverse view perspectives and b) help with mitigating the SfM doming issues. Here’s an example of a curved flight mission - attached. Let me know what you think.