Provide it with max number of frames or time as usual. However, in evaluation mode, it will not test randomly. Rather, it will test exhaustively and record all relevant scores. Furthermore, it will record some more information including a score which will be a generalization test. This score will be calculated by estimating the pose of all other cameras (with sufficiently shared points) in the subset, and calculating the average reprojection error. Also recorded will be the number of other cameras that were able to have their poses estimated, to help prevent skewed results.
If you can get sufficient data (e.g. run this with lots of different sequences, producing many files) then you might be able to more effectively calculate appropriate scorecard values. Later on, you can look at tasks for actually finding better features, and/or doing learning.
Provide it with max number of frames or time as usual. However, in evaluation mode, it will not test randomly. Rather, it will test exhaustively and record all relevant scores. Furthermore, it will record some more information including a score which will be a generalization test. This score will be calculated by estimating the pose of all other cameras (with sufficiently shared points) in the subset, and calculating the average reprojection error. Also recorded will be the number of other cameras that were able to have their poses estimated, to help prevent skewed results.
If you can get sufficient data (e.g. run this with lots of different sequences, producing many files) then you might be able to more effectively calculate appropriate scorecard values. Later on, you can look at tasks for actually finding better features, and/or doing learning.