Are parameter correlation warnings in the Project Status Report after a calibration something that needs further action?

A high correlation between some of the camera parameters does not necessarily mean a problem with the calibration, especially if the correlation is less than 95% and if between a lens distortion parameter and some other parameter.

It is common to see correlation warnings between the P1,P2 (decentering lens distortion) and Xp,Yp (principal point) parameters. For some lenses you will also see them between the K lens distortion parameters. A quick background: A correlation of 100% between two parameters (eg. P1 and Xp) means that the adjustment algorithm cannot decide between adjusting P1 or Xp to get the same result. 100% correlation means that a change in P1 or an equivalent change in Xp gives the same final error – in this way the adjustment algorithm does not know which one to adjust. In the end it does not mean the final solutions for P1 or Xp are wrong, it just means the values are quite uncertain.

Often when parameters are highly correlated we suggest removing one of them from the solution (in the case above remove P1 and let it stay at zero) so the adjuster can choose a good value for the other parameter (Xp in case above). Very high correlations and correlations on many parameters often indicates more fundamental problems with the calibration setup though. You should look at the positions of the cameras – ensuring there are rolled positions (camera rotated to portrait orientation at a few locations) and ensuring strong angles. Also look for good fill of the view finder with the grid.

Correlations between parameters is a fact of life for such adjustment solutions. In fact all parameters are correlated to other parameters to some degree. It is only when parameters are highly correlated that we need to take action. Generally we would not be worried until a correlation is over 95%. In fact in future versions of PhotoModeler we will increase the threshold for that warning to 95%. You can do the same now by opening the Preferences dialog’s Process tab and changing the “Only report parameter correlations over this %:” setting in the Project Status Report section to “95”:

PhotoModeler Pro 5.2.2 and later have an automated method for removing parameters with high deviations and high correlations. This method uses a 95% threshold for correlation. If the automated parameter system can remove a parameter without significantly increasing the total error then it will – otherwise the parameter will remain and the warning will appear again.

These warnings may also appear during a field calibration (‘field calibration’ option during a normal processing). In this case there is no automated method for deciding which parameters are appropriate. You can use the same method as the automated technique though. You remove the parameter that is highly correlated (starting in this order K3, K2, K1, P2, P1, Fw). Ie if P1 and Xp are highly correlated, set P1 to zero and remove it from the adjustment (see Preferences under Process tab). Take note of the total error and rms residual error before and after parameter removal. If it does not change or only increases a small percent then P1 can be removed. If the error increases significantly then P1 should perhaps be added back. It can be a difficult decision sometimes and if you are unsure please do not hesitate to email your pmr file to support@photomodeler.com and we can provide more feedback.

To summarize: a) do not be overly concerned with correlations under 95%, b) if you have parameters correlated over 95% or many parameters with deviation and correlation warnings then consider improving the photography, and c) you can experiment with parameters removal in field calibrated projects by comparing pre and post total error and rms residual.


©Copyright 2017 Eos Systems Inc.