Automatic Timeline  Visualization of Paintings

"Automatic Timeline Visualization of Paintings" (Tom De Smedt, Thomas Crombez & Lucas Nijs, 2016) is an interdisciplinary art & science project that applies techniques of Artificial Intelligence and Machine Learning to Art History. 

 

 

Visual timelines of artists and artworks have played a crucial role in art historiography since the publication of Alfred H. Barr’s catalogue Cubism and Abstract Art in 1936. This book, which accompanied an exhibition by the same name at the New York Museum of Modern Art, prominently featured a diagram of artistic influence on its cover. Since then, timelines have aided students and scholars of art history in identifying watershed moments, or charting the influence of artists and art movements. A notable example is the ‘Great Wall’ of images from Western art history by time period, constructed by British painter David Hockney in 1999. It led Hockney to conclude that the sudden rise of realism around 1420 was to be attributed to new optical techniques, notably the use of concave mirrors.

We propose an automated method for visualizing images from art history. To this end, we have automatically collected a large-scale corpus of artworks by notable painters. The corpus consists of over 3,500 painters mined from Wikipedia, including their name, age, gender, period, style, genre, bio, keywords and impact factor. This information was extracted automatically from Wikipedia articles using text mining techniques and weak supervision. For example, the impact factor is based on the Wikipedia article length and the number of Google Search results for each painter. The corpus has over 5,500 JPEG images (i.e., reproductions of known paintings) of reasonable quality. Each image has a star rating manually assigned by two annotators. Statistical analysis shows that most painters in the corpus have low impact while only a handful have very high impact, and that most painters tend to outlive the average life expectancy in their time period; that female painters have been historically underrepresented; that Amazon reviews focus on painters with a high impact; and that Amazon reviews of art books on notable painters are more often written by highly-educated adult women. 

We are using the corpus to visualize a timeline of the history of painting. If such a wall piece is 10x2m, then it can display about 30% of all paintings in the corpus at 10x10cm. If more interesting paintings are displayed larger, this leaves less room for other paintings. The challenge is to automatically select a subset of paintings and fill the wall in a meaningful way: paintings should be displayed in their time period, paintings in the same style should be clustered, paintings by related authors should also be clustered, and so on. Our current approach is as follows. Keywords for each painter are extracted from the links in the corresponding Wikipedia article. Along with time period and style, these are used as features in an unsupervised k-nearest neighbor classification model (k-NN). It will classify painters in the same period, same style, and sharing more Wikipedia links, as more similar, e.g., Pieter Bruegel the Elder ↔ Pieter Bruegel the Younger. Styles that occur more frequently in the corpus such as Impressionism and Cubism – which may or may not reflect reality – are assigned an anchor position on the wall. Painters with a high impact and paintings with a high star rating are positioned near their respective anchor. Then, similar painters are iteratively positioned near those painters, using spreading activation. This means that the algorithm will start from an anchored painter and radiate outwards to find a free spot. For each painter, we display the JPEG image with the most stars and the highest resolution. 

 

Credits

AUTHOR ROLE AFFILIATION
Tom De Smedt * development EMRG
Thomas Crombez development EMRG
Lucas Nijs development EMRG

References