Taine

A project on image style identification via various approaches from computer vision field.

Contents

0. Data overview
1. Vanilla logistic regression
2. VGG7, Cross-Entropy
3. VGG7 embeddings + PCA + KMeans clustering
4. VGG7 embeddings + PCA + LGBM
5. EfficientNet_B2 pretrained on ImageNet, fine-tuned via TripletLoss
6. EfficientNet_B2 embeddings + PCA + TSNE
7. EfficientNet_B2 embeddings + PCA + KNN
8. Models metrics comparison

Metrics comparison

Cluster separability illustration

Notes

The project is named after Hippolyte Taine — a French historian, critic, and philosopher from the 19th century. In his Philosophie de l'art book he considers an artwork not as a self-contained message, but rather as a representative of a bigger cluster of artist's creations characterized by his style. That cluster, in turn, is a part of a bigger group of artworks sharing a common epoch and region. And further those groups are a part of human art history.