Re: 'Re: Ups.

[email protected]
Newsgroups gmane.comp.gnu.gift.general
Message-ID <[email protected]>
wow. i look away for 1 sec, and all this activity. ;)


On Thu, Apr 13, 2006 at 04:45:31PM +0100, David Squire wrote:
> Wolfgang Müller wrote:
> 
> [snip]
> >>It is the GIFT 
> >>feature extraction that *requires* the rescaling to 256x256.
> >>    
> >
> >Yes, I remember the discussion. Many things the GIFT does are due to 
> >decisions that seemed reasonable at the time. 
> Well, the original source of this decision was that the first image 
> collection we used it on was one from Television Suisse Romande, in 
> which all the images were 256x256. Rescaling to this was a hack to allow 
> other images to be indexed with minimal changes to the code.
> >Makes me think of this article you quoted 
> >at the time that was about the lifetime of engineering decisions.
> >
> >  
> Are you thinking of the one ultimately linking the diameter of the Space 
> Shuttle's booster rockets to the width of a horse :)
> 
> 
> BTW, I do think that not rescaling will cause serious problems, since 
> the feature extraction code is premised on 16x16 blocks at the lowest 
> level, and using images of different sizes will result in the same block 
> feature ID being used to refer to different image locations for 
> different images. IDs are assigned simply by counting.
> 
> Regards,
> 
> David

having looked through the code, the real trick is that gift does multiple 
passes at "quarter scale".
the algorithm should work the same for 256x256 (256->16->4->1) as for 1
024x1024 (1024->256->16->4->1).

having said that, i've been working in that direction, but i'm still 
honoring the origional 256x256 resolution. plus, as i said earlier, 
my feature files, and the files gift "works" from in gift_home are 
bitwise identical.

so no david, i havent broken it. just kicking arround ideas. ;)

Julia Longtin <[email protected]>
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.