[Sigia-l] OT: Library Thing

[Eric Scheid]

On 29/9/05 4:57 PM, "Listera" <listera at rcn.com> wrote:

>> I'm yet to see a machine do a good job of applying subject/depiction
>> meta-data
>> to images, moving or not.
> 
> Then you haven't seen, for example, face recognition software that, even at
> their worst, can do better than a mere human being at his best on a volume
> basis or pattern recognition software that can identify, label and sort
> objects at a rate no homo sapien could hope to match, etc.

Face recognition software - what kind of meta data does that generate? It
produces lots of discrete *measurements*, which it can then compare to a
collection of yet more measurements. It's a stretch to call any of that
metadata, certainly not metadata that users might try searching through,
other than by offering yet another face image for it to draw measurements
from and compare to.

Then there's also the question of fidelity -- I've seen machine processing
of items but they have typically gross/coarse distinctions. Things like
recognising that the shape in front of it's sensors is a horse or an
elephant or a kangaroo ... but not doing so well at distinguishing a rock
wallaby from a big red or some dozen other species. Then there's the tricky
problem of perspective: is the smaller specimen an adult in the distance, or
a joey hopping about. "Joey" is one such metadata term in the APT.

I don't doubt that such things are possible, eventually. I've not seen any
evidence that that day is here.

e.