DRAM Dataset (Diverse Realism in Art Movements)

DRAM dataset is the first dataset to introduce a fully labeled test set for the task of semantic segmentation of art paintings. The dataset uses a subset of 12 classes used in the PascalVoc12 dataset: Bird, Boat, Bottle, Cat, Chair, Cow, Dog,Horse, Sheep, Person, Potted-Plant, and Background.

The dataset consists of 5677 unlabeled and 718 labeled paintings from 152 painters. The dataset is divided into 5 categories: Realism, Impressionism, Post-Impressionism, Expressionism and 'Unseen', each holding paintings from a specific art movement. the Unseen category appears only in the test set and it consists from 135 image of the following art movements: Art-Nouveau, Baroque, Cubism, Divisionism, Fauvism, Chinese Ink and Wash, Japonism and Rococo.

The dataset was constructed for the perceptual task of understanding how computers see images from various styles.

Potential use cases: 1) Evaluating semantic segmentation models on a diverse and complex domain. 2) Investigation on latent representations of abstraction-varying art styles 3) Domain adaptation solutions for semantic segmentation of art paintings.