Re: [help-GIFT] RFC: MRML for adding/removing documents to GIFT

[Wolfgang Müller]

Hi,
> I think that I personnaly prefer
>
>  <mrml session-id="44">
>          <add-document collection-id="TSR500">
>                  <document-list>
>                       <document location="http://some.url/haha.gif"/>
>                  </document-list>
>                </add-document>
>  </mrml>

Cool. Me too, by the way.

> as notation because it is very close to the rest of the mrml
> communication.
> The names are well readable and understandable, that is important.
> Besides the adding and removing of images, there are a couple of other
> commands that might be needed:
> - create a new image database,
> - delete an existing database,

Oh, yes, I wrote that in a reply to my initial message, as I see now this 
response did not go to the list, but only to myself :-(. Here I paste it:

quote:

...and of course there is the add user/grant rights part of the question...

For granting rights to modify etc. collections, I would be for mapping SQLs 
or mySQLs behaviour on some MRML tags.

For granting rights to access documents, i'd be for a unix-alike 
document-wise rights system, probably with the following classes

user,
group,
others (local),
group(remote).
others(remote).

(snip)
unquote

> - get all image URLs in a certain database (is this already possible?).

Indirectly, yes. Give a random query for 2000000000 elements. 

>
> This will of course mean that a secured (encrypted) communication is
> desirable and that user management with password is necessary!
>
Anyone who extends the server by an *optional* ssl is very welcome. Also 
welcome are people writing a ssl tunnel in perl. The tunnel could use unix 
sockets for communication with the main server.

>
> Once this is in place, there are of course lots of interesting
> questions, such as choosing a feature set for a new database to be
> created, selecting the rights of users (or user groups) on the creation
> of a database.
>
I agree with you there.

> Somebody with good database knowledge should be a great
> help with this as these should all be well known problems.

Yes.

Concerning wour remark about the use of features when creating a database, I 
guess it's best to use query-paradigms for that. A feature extractor should 
know what query-paradigm it can extract features for, and the collection 
describes what query-paradigm it contains indexes for. The algorithms know 
which query paradigms they can query.

Till now, vanilla MRML avoided more or less explicitly communication about 
the internal representation of things. So also in this case, I would rather 
go for solutions which keep things like information about which features are 
used, as far as possible implicit. IMHO, using query paradigms seems to be 
fit for the most urgent needs. 

However, concrete suggestions for an extension of this mechanism would be 
welcome.

One thing that might be interesting concerning adding new documents (and that 
is not about representation), would be to indicate the speed/precision 
tradeoff when adding images by a fraction. Like 
speed-performance-quotient="0.1" (please suggest something more intuitive, if 
you find it, another idea would be to fix the 
maximum-extraction-time-seconds="2"). This would enable each indexer to make 
decisions like: "I have time for three wavelet scales and a small color 
histogram", without being fixed on a given query method and data 
representation.

Cheers,
Wolfgang