26 Nov 2015 Pixolution Testing Keyword Suggestion Services
The visual computing group at HTW Berlin, under the supervision of Prof. Kai-Uwe Barthel, who is also behind the technology of Pixolution and Picsbuffet, have recently launched Akiwi, a semi-automatic keyword suggestion tool.
It’s been out for a while already, but they’ve just released a new and improved version. It is intended to overcome the inaccuracy of fully automated keywording by adding a touch of user interaction. And it’s pretty sweet!
How it Works
The Akiwi website is an open demo of the technology. It could be plugged into any agency, collection or DAM system. The current website uses 22 million images from Fotolia.
By uploading an image, it finds visually similar images from the library and collects the keywords from those matched images. Accuracy can be improved by simply clicking on any of the matched images that are closest to the uploaded sample.
Clarifai has a tool that does the same task, but works differently. It analyses the image itself using machine learning technology and generates the keywords from a fixed list.
The Akiwi method requires a library of already-keyworded images to function, but this can be an advantage, providing the flexibility of working with any controlled vocabulary or unstructured keywords.
Use in Microstock
This functionality will soon be available in the commercial product ‘Pixolution Flow’ enabling to work with their own image sets and specific keyword vocabulary. Microstock agencies could use it to offer a keyword suggestion service to contributors similar to that introduced by Shutterstock in 2013.
The Shutterstock tool is more rudimentary and much less automatic than Akiwi: contributors must surf through images, select similar images to the one they want to keyword, then go through the list of keywords and finally select the relevant ones to be included in their image. Akiwi achieves the same thing with much less user interaction.
Of course very few microstock contributors do their keywording on agency websites, so this tool would be more useful for microstockers as a plugin to Lightroom or similar applications. It could also be a nice add-on for services like picWorkflow, or as a tool to accelerate the keywording process for service providers like picWorkflow again (which also offers a keywording service) and the recently launched Kboarding.
Here’s a demonstration video from Dr Barthel himself showing how it works:
Bob Davies
Posted at 18:57h, 26 NovemberIs this really any different to the 5-6 other sites that sites that do this? I’ve tested a few ways of suggesting keywords like this over the years, but always comes out with very noisy (close to, or well within the ‘spam’ range) dataset.
Starting with a fully populated set always gives a worse set of keywords than starting with an empty set and only adding those which are relevant.
Maybe I’ll reconsider when any of these pseudo-algorithms can favour relevance over noise. Until then I’ll stick with humans 🙂
Kai Barthel
Posted at 08:34h, 27 NovemberBob, I totally agree that automatic image tagging does not always work.. akiwi is not intended to work fully automatically.
The main difference between akiwi and other automatic image tagging systems is the fact, that you can get the keywords you want pretty quickly by using it like this:
1. Submit your photo
2. Click the image which is very similar to yours. If there are no similar images click or enter the keyword that best describes your image.
3. Then click finalize to clean up the list of proposed keywords (add and/or erase keywords)
We have made user tests and found out that compared to manual tagging, working with akiwi (as described above) resulted in a speedup factor of 5 – 10 depending on the images.
Bob Davies
Posted at 16:52h, 27 NovemberDon’t get me wrong, it’s a pretty impressive tool for speed. Sadly I’ve never seen any of these tools give as good quality a set of keywords as a starting-from-scratch manual. Speed is fine, but speed isn’t what sells images, quality is 🙂 Top-down can come close, and has always been able to hit that max-50 quickly, but that’s really not the point… sometimes 8 keywords are enough, if they’re the right 8 🙂
Bob Davies
Posted at 16:53h, 27 NovemberThough I am biased I admit. I spent almost 5 years trying to perfect the keyword suggestion algorithm and barely even came close 🙂