...your personal museum tour guide...
An intelligent recommendation system for paintings using image recognition
Museums are big business: in Canada alone they generate 2.5B $/year, employ 37.000 people and attract 54m visitors/year
However, recent studies have shown that the museum market has a alot of trouble both attracting a younger demographic as well as building a loyal customer basis...
Enter Muse: a modern and intuitive way to discover and navigate museums using state of the art AI.
To mimic a toy museum, 500 images accors 5 different artistic style were scared using the Google API
To create an image recommendation system, we used transfer learning on the VGG-19 dataset by removing the classifying layer as well as the last convolutional layer and simply used those pre-trained weights to obtain and embedding
Next, we used K-NN to obtain the closest recommendations using cosine similarity
Validating a cold start problem is typicaly quite hard, so instead we decided to ask Muse another problem: can it detect similarity of styles. It turns out that it can with ~80% accuracy. A T-sne plot shows how well Muse clusters different styles
- deployed as webapp using Flask on Heroku
- deployed as php code on louisdethanhoffer.com
- Release MVP