Style or Signature? Artist-Disjoint Evaluation of Style Classification in Frozen Vision Embeddings
The paper says style classifiers lose ground when they cannot lean on the same artists appearing in training and test data.
Rory Ashton tests frozen vision embeddings on 320 paintings across four twentieth-century movements, using an artist-disjoint setup that holds out each painter. Under that protocol, 5-NN style accuracy drops from 0.87 to 0.77. The fall is uneven: Impressionism and Cubism are largely stable, while Surrealism drops by twenty points. The result holds across four encoders, leading the paper to argue that artist-disjoint evaluation is needed to tell style recognition from artist recognition. ArXiv · AI/CL/LG's note
Rory Ashton tests frozen vision embeddings on 320 paintings across four twentieth-century movements, using an artist-disjoint setup that holds out each painter. Under that protocol, 5-NN style accuracy drops from 0.87 to 0.77. The fall is uneven: Impressionism and Cubism are largely stable, while Surrealism drops by twenty points. The result holds across four encoders, leading the paper to argue that artist-disjoint evaluation is needed to tell style recognition from artist recognition. ArXiv · AI/CL/LG's note
score 4