Quilt

Can a Machine detect Interestingness


Can a machine detect the interestingness of a picture? I wanted to see if it could be done and I also wanted to change the way we view thumbnails at the same time. Thumbnails are normally scaled down images made to fit into a generic box. I dislike this, its done and boring. I am trying out a new way of cropping interestingness - to make somehting more enticing to click as you cant see the whole image, not to make the experience quicker, to make it more fun. 

I loved playing guess the object as a child and I suppose this is the very same thing, albeit slightly automated.

Take this image for example:

It renders the following image preview.

According to the transforms this is the most interesting area of the image, it is interesting how it is the refleciton that is more interesting as it has the added element of the interference of the water, the reflection is the most visually interesting part of this image, the filter is not always right but generally crops an interesting area of the image. With some tuning this algorithm could be further tweaked.

I calculate "interestingness" by inspecting each pixel with the neighbouring pixels and figuring out how different it is from the rest, when you do this for every pixel, it creates a black and white image of contrast thresholds. 

 

This is then aggregated to a rigid map which then translates "interestingness" coordinates to crop the image. Its fairly obvious where the interesting part of this image is.

Thought I might explain how I acheived this, it may not seem like there is anything going on but there is a lot under the hood :). This rendering happens every time you upload an image - I am not a real developer, I built this site for fun and to revitalise my out of date coding abilities and to make an open sharing platform for anyone to use. I suspect there may be some bugs but  Iwill iron those out as we go. 

 

 

no url provided


Help Support this site by buying some Liqueurs
by: User Number 1
dateStamp: 2013-11-25 17:07:44
Category: Design



Meta