[ExI] Fashion

Anders Sandberg anders at aleph.se
Sat May 18 15:16:49 UTC 2013


On 18/05/2013 14:57, spike wrote:
>
> >..."Dominant shirt colour must be present in tie pattern"
>
> It doesn't actually define the term dominant shirt. Extrapolating from 
> the definition of dominant term in a polynomial, they mean your 
> largest shirt.  OK so I know what color that one is, but I don't wear 
> it often because it doesn't fit well.  I would think they would 
> restrict their interest to whatever shirt or shirts one is wearing at 
> the time, perhaps the outermost shirt, even if that one isn't actually 
> the largest.
>

Hehe. Operationalizing "dominant shirt color" is also tricky: it could 
be the color found on the largest area, but some patterns have several 
similar colors that add together. And "be present" clearly has a degree 
of fuzziness. Presumably one could handle this by doing clustering in a 
suitable color-space for shirt and tie, select the shirt color cluster 
with the most weight, and then ensure that the overlap with one of the 
tie clusters is above some threshold.

> >..."Bonus/penalty if tie is darker/lighter than shirt"
>
> Excellent suggestion.  I never thought about it, but my bride 
> apparently has, for she is the one who buys my ties.  I do tend to 
> grab one at random.
>

Note that "darker" is also a slightly tricky concept, since it is based 
on human subjective perception. The way to automate this is to use a 
spectrocolorimeter to measure tristimulus values and then multiply by 
the human sensitivity function - blue light appears darker than the same 
intensity of yellow light. Normal cameras likely can do the same job 
with slightly less reliability.


>
> >..."Correlation length of shirt pattern should be dissimilar to 
> correlation length of tie pattern"
>
> This suggestion needs further explanation in terms of correlation 
> coefficient.s
>

You calculate the autocorrelation function (how much the pattern 
correlates with itself when translated by a certain distance) and then 
look for the characteristic length - either the decay constant if the 
pattern is random, or the distance to the first peak for a repeated 
pattern. This is a standard operation in signal processing and computer 
vision, and can be done quickly in Fourier space. See 
http://www.ijcte.org/show-46-593-1.html for an example of how to use 
this on fabrics.


>
> >..."Width of tie should be within X% of shirt collar and jacket lapel" 
> Anders Sandberg
>
> We would need to define where to measure the width.  I would assume 
> maximum width of all given parameters without further information.
>

Yes, and during different periods ties have had different divergence 
angles. Still, most of these measurements are very loose - 10% errors 
are likely quite acceptable. It is not rocket science :-)


> Our computers and phones have cameras in them now.  It seems like we 
> should be able to stand in front of it and have it use some set of 
> criteria, then tell us 'good to go' or 'lose that revolting charteuse 
> tie,' etc.  We could show it all our clothing and let it make 
> suggestions depending on the occasion, which it knows too, since the 
> occasion is written in the calendar app.  It would be Silicon Eye for 
> the Straight Guy.
>

Love that name.

Given current research, it might not be entirely impossible to make. 
Segmenting clothing from pictures seems very doable: 
http://cs229.stanford.edu/proj2009/McDanielsWorsley.pdf

And this paper suggests where to buy similar clothing as a person wears 
in a photo:
http://labs.yahoo.com/publication/getting-the-look-automatic-clothing-suggestions-for-everyday-photos/ 
( paper at http://tinyurl.com/d67mj5n ). If you use this system on a 
collection of pictures certified to be of well-dressed people, 
presumably it would suggest what to get.

Ooh! Just found this paper: "Hi, Magic Closet, Tell Me What to Wear!" 
http://www.lv-nus.org/papers/2012/magic_closet-MM12.pdf
Wow. It seems to do almost all you wish for. I especially like that they 
use latent variable mining of example clothes: those formal rules we 
started with are just the tip of the style iceberg, and this way the 
system can in principle learn implicit and unwritten rules, even rules 
no style guru has ever articulated.



-- 
Anders Sandberg,
Future of Humanity Institute
Philosophy Faculty of Oxford University

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20130518/51e567a1/attachment.html>


More information about the extropy-chat mailing list