This is triggered by a question about where to share photos, since my all-but-open-sourced project is about targeting photos to the viewer’s needs in the moment.
Around 2000, having a pile of photos I considered art,* I realized that the effort choosing ‘the best’ felt wasted, even if I were to reach Ansel Adams’ name recognition. I don’t want 1-5 photos to stick in people’s minds to keep me famous for a century or three.
I decided to build a photo-based ‘brain’ that would pick photos dynamically in dialog with the viewer,** and wound up displaying the photos in pair format. I started pairing photos by matching color properties and keywords, reaching up to 20% accuracy in predicting interesting pairs. Then simple (and lame) neural nets jumped accuracy to 65%, as recorded here:
http://phobrain.com/pr/home/schedulaid.html
Now accuracy is around 90% using raw imagenet model vectors, if I allow the same pics to be in both training and test pairs. (It works well enough to forget for now.)
Screenshots are of the workbench for browsing unseen pairs and labeling them accept/unaccept (a/u in screenshots). The other options generate pairs randomly, by color matching, and by various neural net approaches including the personally-trained ones.
If any programmers want to be the first to try it so I/we can smooth the setup phase for civilians, I can wrap up open sourcing (AGPL). It’s in Java (web server and tools) and Python (ML and tools).
Thanks,
Bill
- (Ivan Karp of OK Harris: “What you have here is fine art photography.”)
** Applying tech learned while doing quality assurance on ad servers.
*^N My first app when they replaced the card reader interface with terminals at the city college, ~1985: