This project was created for http://school-slon.ru/ hackathon. As part of the assignment model for determining time period painting is belonged was designed.
Firtly we downloaded a large amount of data with paintings from
https://www.wga.hu
For better class balance we decided to unite classes beloged to 201-1300 гг
For solving task this architecture was chosen:
- Vgg19 network up to conv_5_1 (1)
- Applying gram matrix (2) transformation for (1)
- Filtering 8000 features from (2)
- SVC with 13 classes
On test set we've got sush results:
- MSE: 4 (it means the model wrong on one century in average)
- F1-score: 0.5
Confusion matrix:
