Re: [help-GIFT] alive?

[risc]

On Tue, Apr 18, 2006 at 10:55:35AM +0100, David Squire wrote:
> address@hidden wrote:
> >Julia,
> >
> >That's great. You're motivated, you're pushy :-) . Given that the output 
> >is equal to the previous output, I guess the main part of your sources 
> >should be in GIFT quite soon. Without having looked more at your source 
> >than before, I can tell you the questions I see:
> >
> >1) I did not quite understand the thingy about the "aggregator"?
> >2) Are you able to write "feature information files"?
> >3) Given that Image Magick is already a prerequisite for successful GIFT 
> >installation, the current state of your code introduces a new dependency, 
> >right? Probably everyone has it anyway, but I still think it would be best 
> >to move to the use of ImageMagick as file reading library. If you've got 
> >the time, it would be worthwile adding that, and it would be very much in 
> >your own interest, as you would save the convert call that has to be 
> >issued for each image.
> >
> >David, Julia: sounds reasonable?
> >  
> 
> Yes, that does sound reasonable to me.
> 
> There is a broader point too though. I would like to encourage the idea 
> that the feature extraction part of the GIFT should be in some way 
> pluggable. When adding a collection, the user should be able to specify 
> which feature extraction system he/she wants to use, and there could be 
> multiple ones shipped with the GIFT, or others could be available from 
> third party sites.

sounds great to me. porting my code to imagemagick as we speak.

a pluggable system might be in order BEFORE we make a release, as my code 
performs 2/9ths WORSE on a non-sse2 enabled machine.

> 
> I guess that this would mean that (at least) an attribute would have to 
> be added to the collection element in the gift-config.mrml file so that 
> the system could be aware of whether features are comensurate for 
> different collections.
> 

personally, i'd love to see some documentation on the feature file format, 
and a "clean reference implimentation". i've overhauled your code, and i 
still cant make heads or tails out of some of it. comments would be nice. ;)

> I am fairly sure there was a discussion about this a few years ago, 
> either on this list, or via email - I will have a look for it.
> 
> This would make it easier for various GIFT users to share their 
> specializations of the system (e.g. I have a version that uses 
> annotations too, and I know that Henning has done stuff using a greater 
> number of greys, more texture orientations, etc.)
> 

i'd be interested in seeing some of this work as well, as my input images 
are all greyscale.

> Regards,
> 
> David
> 

to you, 

Julia Longtin <address@hidden>