Different Results from Same Data

I am trying to understand the processing of data.

I processed the exact same set of images, with the exact same settings, several times and get different elevations each time.  My output is set in feet, NAD83(HARN) / Tennessee (ftUS) and set Geoid Height above GRS_1980 to -400 ft.

I found a very distinct, single tie point that is similar in all processings.  After processing three times, the elevations of the point varied (311.92, 325.59, and 334.28).  Computed camera elevations varied too (490.79, 490.81, and 491.09).  Using Phantom 4 with image EXIFs, without GCPs.

Why do I get differing results from the same data?

Thank you.

Dear Matt, 

We would need to further investigate to understand where the difference comes from. 
More specifically, we would need a Quality Report and Log file of the project. If possible, one Quality Report and Log file for each of the three processings you have done. This will give us valuable information to understand what could be improved. 

Please attach the documents when submitting a request here: https://support.pix4d.com/hc/en-us/articles/hc/en-us/requests/new

Thank you

So in the end, what was the cause of differing results ? Any news ?


Hi Joe, thanks for following up. 
Matt sent us the Quality Reports of his two projects.

The project has 21 nadir images over a parking lot that had a few trees and cars. The Processing Options Template was correctly selected as 3D Maps and the camera model was selected from the Pix4D camera database (see the icon on the left): 

The difference came from the Camera Optimization that was different in both runs of the projects (17.6% and 6.36%), see below: 

Project 1:

Project 2: 

We recommend to have the Camera Optimization below 5% , if the difference is larger, there could be issues in the reconstruction. Note that this is the general case for cameras from the Pix4D camera database. A few options to improve this in the projects above:

Take more images

We consider a project with less than 30 images to be very small. When there is not enough information, it can lead to different results during the calibration of the images. 

Therefore, we recommend to take more images, which should improve the calibration of the cameras during step 1. Initial Processing. 

(optional) Linear Rolling Shutter correction

As seen above, the project is using a DJI camera. Most DJI cameras have a linear rolling shutter on their camera, which can lead to the Rolling Shutter Effect depending on the drone speed (the faster it flies, the more probable it is that there is a Rolling Shutter Effect). A visible effect is that the reconstruction is a bit curved. This can be corrected in the software as described in this article: 

All Prior option
During 1. Initial Processing, the software tries to optimize the camera model. As there were not many images it found different solutions each time. As the camera is from our camera database, the model of the camera is supposed to provide good initial parameters. Therefore, we can tell the software to “trust” these parameters and to remain close to these values for the optimized parameters.

This can be done by selecting the option “All Prior” in Process > Processing Options > 1. Initial Processing > Calibration > Internal Parameters Optimization. Then, step 1. Initial Processing should be reprocessed.