Very interesting, though I would love to see more data in there. 3840 is such a small number for what is actually out there.
I noticed that a lot of the automated stuff, like glasses, is getting a pretty high false positive rate. ~25% of the 'glasses' ones are actually no glasses. Would be very cool to see more out of this though.
Also, don't they need to attribute the image back to where they got it on instagram to comply with the API usage terms?
According to [1], they collected half a million photos from instagram, selected 20,000-30,000 photos from each city from that, then filtered the set down by using Mechanical Turk.
I noticed that a lot of the automated stuff, like glasses, is getting a pretty high false positive rate. ~25% of the 'glasses' ones are actually no glasses. Would be very cool to see more out of this though.
Also, don't they need to attribute the image back to where they got it on instagram to comply with the API usage terms?