11 comments

[ 3.0 ms ] story [ 38.1 ms ] thread
I've written the cropping algorithm being used at Facebook, here is a write-up of the method if you are interested: http://blog.vjeux.com/2012/facebook/best-cropping-position.h...

I tried to use saliency-based approaches but they don't really work that well for non artistic photos. For example, this image[1] would have edges everywhere but next to the faces.

If you have some time, I would be interested in seeing the results of your approach in my example images[2].

[1] - https://fbcdn-sphotos-c-a.akamaihd.net/hphotos-ak-snc6/p206x...

[2] - http://blog.vjeux.com/wp-content/uploads/2012/10/cropping.ht...

Very interesting approach thanks, unfortunately we don't have the social features ...
From my photography experience which i also apply to graphic design the Rule on thirds and the Golden ratio are the best ways to locate / create interesting parts of an image. This can be achieved by clever cropping too.

http://en.wikipedia.org/wiki/Rule_of_thirds http://en.wikipedia.org/wiki/Golden_ratio#Aesthetics

Doesn't this rely on the photographer having framed the photo with that in mind originally? Given a set of arbitrary photos I'm not sure the rule of thirds could be applied in any meaningful way to pick the "interesting bits".
Reddit's thumbnailing script works nice as well, https://github.com/reddit/reddit/blob/master/r2/r2/lib/scrap...

edit: I even think you are using a similar method.

Yes I'm using a similar method in the first section of the article (information theoretic solution), but I think that reddit is using the entropy of an image instead of the self-information. The self-information yields to better results in our case (empirically). In the latter sections I use different methods.
(comment deleted)
Sounds VERY interesting. I was wondering - is there a python implementation of this algorithm somewhere?