If you are solving for more than just focal length in an Inverse Camera project it is important to check the Project Status Report after a successful processing. You are looking for high correlations between the camera parameters and high standard deviations. Correlation values over 95% are suspect and if a camera parameter being solved by inverse camera has a correlation over 98% you should probably not be solving for it (you should instead leave the parameters at a reasonable default, such as square pixels and centered principal point).
In this example, from a real-world single photo inverse camera project, part of the status report has these values:
Photo 1 Standard Deviations [and correlations over 80.0%]
Omega: 0.016 rads [ Phi:97.1% Fw:97.9% ]
Phi: 0.023 rads [ Omega:97.1% Focal:83.6% Fw:99.1% ]
Kappa: 0.013 rads [ Z:91.9% ]
X: 0.216 m
Y: 0.290 m
Z: 0.423 m [ Kappa:91.9% ]
Focal: 0.250 mm [ Phi:83.6% Fw:81.4% ]
Fw: 0.195 mm [ Omega:97.9% Phi:99.1% Focal:81.4% ]
You can see a correlation between Fw (this is 'format width' and what is computed if you select "format aspect" in the Inverse Camera section of the photo's property dialog) and the station angle phi of 99.1%. This means the solver could not distinguish errors in Fw versus errors in Phi. This means that neither have solved well. Also note the one-sigma standard deviation of 0.195mm for Fw as this somewhat high because Fw error can introduce quite a bit of error into a project.
This is the same project solved with just a focal length inverse camera setting:
Photo 1 Standard Deviations [and correlations over 80.0%]
Omega: 0.003 rads
Phi: 0.003 rads
Kappa: 0.004 rads
X: 0.207 m [ Z:92.3% Focal:97.4% ]
Y: 0.270 m
Z: 0.139 m [ X:92.3% Focal:91.6% ]
Focal: 0.146 mm [ X:97.4% Z:91.6% ]
Note that we still have quite a high correlation between focal length and the camera station position but often this cannot be avoided as it is difficult to make reasonable assumptions about the true focal length. Also note that the one-sigma standard deviations are generally smaller on all parameters so this is a better solution than the one above. In this particular forensic project, other data was checked after the processing (residuals, check distances, etc.) and the results were shown to be reasonably accurate.