I disagree fundamentally with the assumption that icons exist to highlight the “important” bits of the UI.
First off, nine times out of ten, if you have unimportant UI bits you’ve screwed up. Second, if an important bit of UI needs “highlighting” you’ve almost certainly screwed up.
Icons are not workarounds, they are the key to the whole visual metaphor. Several papers have discussed the power of pairing a visual and a lexical stimuli for increasing accuracy and response time in a number of tasks.
Give someone a set of words, and ask them to select the one of them according to a criteria (Like picking the menu item which will save the file from the list of menu options), and it will take them several milliseconds to find it.
Give them a set of images with the same criteria (“Find the toolbar button that saves the document”) and they will take several milliseconds.
Pair the two together (image plus text) and it will take less time to select the menu item.
This makes it pretty clear that eliminating icons throughout the system is a net loss. What is still ambiguous, from a scientific perspective, is what happens in the mixed case – in other words, what is the performance change for the case where “important” bits of the UI have icons and “unimportant” bits do not.
I strongly suspect that performance of selecting a specific “important” item will go up marginally at best (and I could make a strong argument that would theorize performance would plummet, in most cases) and that performance will be so much worse for selecting “unimportant” elements of the UI, that the effect would be a net loss.
I’d be happy to test this out, however, using Gtk+ specifically, if one of you happens to have a grant?