I know it's been a long time since this thread. But just to follow up:
In general, I presume @imagejan has carefully reviewed the TWS documentation which explains all those options, and has used it enough to have an intuition for which features are suitable in different scenarios.
The docs are pretty math-focused and less intuition-based, but you can probably get a feel for how the classification is working by trying different combinations of features with the "Train classifier" button on a small image.
I fuzzily recollect that @iarganda eventually plans/hopes to make some kind of feature browser, so that you can preview the training results with different combinations of features. In general, the combinations are exponential, but it would be doable to at least show a preview of a few commonly useful combinations on a small window of data.
See the TWS documentation on macros for instructions.