I explored the Cooper Hewitt database that contains over 200,000 images and found 2 image sets I was interested in: posters and playing cards. I decided to work with playing cards. This dataset contained around 2000 images of playing cards from across the 20th century, acquired between 1909 and 1986.
Running a t-SNE on the cards produced the image shown below. It revealed 10-20 near complete sets of card and many incomplete sets and unique ones. The cards range from having purely text to more traditional ones having a number and suit to cards that have images like photographs, faces, flags and scenes.
t-SNE output of a set of 2000 playing cards
I was particularly interested in face cards and so ran a face detector on all of them using a Haar Classifier in ofxCv. I also clipped areas around faces from all the cards in openFrameworks.
t-SNE output of clipped faces
My final visualization was made in Processing. On keyboard input, the program displays a hand of 5 cards picked at random from the set of identified face cards. Since the face detector is not completely accurate, the 5 cards in a hand are mostly but not always face cards. Additionally, the clipped faces are run through in a small window that shows both the correctly identified faces and false positives.
Links: