In the current version of PhotoScan (1.3.1), the intrinsic parameters considered in the camera model must be selected by the user in the “Optimize Camera Alignment” dialog. If we aim for best results, this selection is very difficult for the user because there is no statistically valid criterion available for the selection. “Adaptive camera model fitting” option in the previous stage (“Align Photos”) may give a hint, but it often removes too many intrinsic parameters.
RMS reprojection error for the “Control points” and “Check points” (appearing in Reference pane) as well as the sparse point cloud (checked by Show Info command for the chunk) is not a valid criterion, because the pixel coordinates of all of these points are used in the optimization. RMS estimation error of the real coordinates of “Control points,” appearing in Reference pane, is also not a valid criterion for the similar reason.
In statistical terms, these points are “training data” to which the model is fitted. For valid evaluation of the model quality, we need independent “test data.” Otherwise, the model will suffer from overfitting of the model: good fitting to the training data and poor reprojection of the test data.
Therefore, I propose a command as follows:
1. Random splitting of the tie points (sparse point cloud) into the training and test points (in the ratio about 90%:10%).
2. Optimization (minimization of RMS reprojection error) using training points only.
3. Evaluation of the RMS reprojection error for the test points on the basis of the estimated camera intrinsic/extrinsic parameters and point 3D coordinates.
4. Repetition of the above steps 10 times or more to formulate a cross validation.
Hopefully, the command automatically repeats this procedure for various combinations of the intrinsic parameters considered, to find the best combination.
This is a statistically valid selection procedure of intrinsic parameters. It is similar to the selection of explanatory variables in linear regression. Because it would be computationally expensive to try all possible combinations of the intrinsic parameters, I think “forward selection” starting from only “f” and increasing the parameter one by one is practical.
I am happy if the Agisoft considers the implementation of the command like above.
Topic: Command for statistical best selection of intrinsic parameters (Read 4795 times)
