> A hierarchy where each node fits on the screen is perfect for a tablet as
> one can literally remember finger movements to get to the item.
and utterly fails when the collection grows.
the vast majority of people absolutely suck at hierarchical navigation, just as the vast majority of people absolutely suck at remembering numbers longer than a handful of digits.
> I do not see how tags provide such navigation stability.
by removing the need to navigate a non-spatial data set spatially.
how often do people lose their keys?
how often do they forget what their keys look like, or which keys are on their keyring?
> For big collection of images the story is not so good even if one just
> wants to see them.
the movements to tag a big collection of images into the same organizational array as represented by a bunch of folders is approximately equivalent (can be slightly better, can be slightly worse, can be the same; depends on the exact structure).
the problem most people have with tags (me included, btw) is that the mechanisms provided to tag things have SUCKED (as in: too labor intensive, not part of the primary workflow so always extra work rather than "for free" during usage) only to have really weak tools built around the semantic data sets such as "a tag cloud" or "if you show this sidebar and then change the view appropriately, you can reduce the current display by a given tag." which means the input effort totally outweighs the output benefit.
i have, actually, used Plasma Active to tag large sets of photos such as ones i took on a vacation last year. the one thing i really wanted available was a "show me pictures i haven't tagged yet" (something like a negative tag filter) .. it's on my todo in any case :) we have a full "importing large photo sets" type workflow already designed out, it's a matter of getting to the implementation (which will happen eventually; we're moving at a good pace as can be seen over the short lifespan of Active so far :)
keeping in mind that tags are but one aspect of the semantic data we provide access to (time, usage patterns and in-file metadata also are utilized; free text searching, timelines, etc are also provided built into the workflow everywhere) and that the UI limits the input effort and maximizes the output benefit, with a worst case usually being "equal to a hierarchical system" .. it works rather well.
now, one use case i find it is still not good enough for is when you start with a disk full of metadata free pictures that have few if any relationships with each other (iow: migrating to the system from a historically data poor system). we have some thoughts on how to ease such transitions (including mining the existing folder hierarchy if it exists) and will be improving that aspect of things. of course, we first needed a working metadata centric system to migrate too ;) and for touch devices right now, the common use case does not currently include "migrating all the content from my laptop with a 15 year old file structure on it" (which would be my case :) but rather tends to be a "starting from a blank device, adding content to it incrementally during use"