I really think this needs more discussion.

As with all least squares analysis, the weight - the error estimates - for any and all measurements - the sources of error - need to be correct in order for the results, and just as importantly, the statistics quantifying the results to be correct.

I do not believe that a 4 pixel error is a valid estimate for the tie points image matching accuracy.

We can have a long discussion about what it should, or rather could be, depending on the matching algorithms used, but it is quite common to expect sub-pixel matching - sub-pixel error - even as low as 0.1 to 0.2 of a pixel of error. In PhotoScan Pro with minimal optimization and gradual selection removal of the worst points, a point variance RMSE of less than 0.5 pixel is very easy to obtain. So, a 4 pixel error estimate is very pessimistic. An unreasonably pessimistic error estimate.

It also seems to me that entering a much more realistic estimate - in the sub-pixel range - does not in fact change the statistics as much as it should. It still seems that control, whether it is true ground control or camera position values, are given too much weight not necessarily honoring the entered values.

Tom