Figure 1: Samples from WikiART Face Image.
Recent progress in computer vision proves that large generative models can model natural images with a decent quality (BigGAN, GLOW, etc). Nevertheless, there are still vast room to improve the generalization ability and the quality of the generated images, so the community is keen to develop new approaches for better image modeling. As an art fun rather than a machine learning researcher, I have a question “Can we model artworks with the generative models?”, and if so, I believe those models should give us different perspectives of artworks. This is my core motivation of this project to have a general art image dataset to facilitate research in computational art. In this project, I introduce two art image datasets WikiART Face & WikiART General. All the original artwork images are taken from WikiART, an online visual art encyclopedia, with wikiart-crawler, a python-library also developed within this project. For further information including the link to each dataset and basic usage, see the official project’s github repository.
Figure 2: Pipeline to produce single WikiART Face image.
WikiART Face is a dataset of face images produced on top of portrait images (see Figure 1 for sample images). Inspired by one of the largest human-face image set CelebA, all the portraits from WikiART are processed through the pipeline described in Figure 2, that results in 6,095 images across various art movements. Following figures present samples from some art movements.
Figure 3: Samples from 'Ecele de Paris'.
Figure 4: Samples from 'Impressionism'.
Figure 5: Samples from 'Pre-Raphaelite Brotherhood'.
Figure 6: Samples from 'Rococo'.
Table 1 shows the data size breakdown per each art movement.
|Art Movements||Image Size|
|Ecole de Paris||228|
|Naive Art (Primitivism)||72|
Figure 7: Samples from WikiART General.
WikiART General is a dataset of general artwork images taken from WikiART. Unlike the WikiART Face, this dataset contains various type of art from portrait, landscape to abstract paintings (see Figure 7 for some samples). Note that even the portraits are not as same as in the WikiART Face, since they are raw image without face-centric pre-processing (the pipeline shown in Figure 2). This dataset contains 25,612 images in total, where 6,126 in portrait, 5,872 in landscape, and 13,614 in others. Visit the project’s github repository to see a complete statistics.