30 Jun 2015 Picsbuffet – Images Arranged by Similarity on a Zoomable 2D Map

Picsbuffet is the latest work of Dr. Kai Barthel of HTW Berlin and Chief Visionary Officer (get it?) of Pixolution, a company providing technology and research around visual search solutions in the stock photo industry, and others.

It is a “Google Maps style” image search platform for stock photos, utilising Pixolution’s visual search technologies to show a 2D ‘map’ of images arranged by relevance – a fused measure of keywords and visual similarity.

It contains one million images sourced from Fotolia – still the most open and functional API in the market.  Kai says they “could and will” add more images to their database, once they’ve locked down a server solution powerful enough to efficiently sort large volumes of files.  

The relevancy algorithm does a great job of clustering related topics side by side, but you can at times find some ‘hard edges’ of unrelated images.  Asked about this, Kai explained this is due to the quantity of relationships between different topics.   Preserving all image relationships is impossible on a 2D map, but he’s working on changing the format to a ‘graph’ of images to address this issue.

Picsbuffet is an academic experiment, though with obvious commercial implications for any business that has visual content, including product images from manufacturers such as jewellers and fashion labels.  And, of course, stock photo agencies.

Check it out and let me know what you think in the comments.

2 Comments
  • Bob Davies
    Posted at 14:15h, 30 June Reply

    Nice idea, but it looks more like a dev who was focused only on how to render that many images to a HTML5 canvas, not why they would do it. Can’t see it adding anything useful to a buyer’s search experience.

    https://www.dropbox.com/s/fej7pvkhalyvs4v/weird-pic-boundary.png?dl=0
    The boundaries on the grid are weird (seems to have no relationship, just zoned together to make a grid, regardless of whether those images should be in that ‘zone’. I’m guessing here of course but I suspect they’ve been zoned arbitrarily, rather than algorithmically (or at least needs a better algorithm).

    https://www.dropbox.com/s/bzrl4y4xghrncsa/nice-idea-crap-results.png?dl=0
    The search results are also not very clear (not sure if this is just down to Fotolia’s API, or some other part of the render-tree) so probably less useful (or at least more distracting) to users.

    I do like the heatmap, and the concept has potential I’m sure, but there are better ways to do this technically (looks like a pretty standard canvas output on the client-side, which also doesn’t lazy-load very well. Maybe would benefit from some extra web-gl shaders to give some focus or weight to the most relevant (or actively in-focus images)), and better ways to present it to the user (nobody wants millions of search results, they want the 20 best images that match what they entered, even if they entered some search that actually means something else).

    Imho: Nice idea, needs more work 🙂

  • Kai Barthel
    Posted at 16:44h, 30 June Reply

    Currently picsbuffet is only supported by the desktop browsers Chrome and Opera .
    A short explanatory video can be found at: https://www.youtube.com/watch?v=D7JUo_IEDpk

Post A Comment