Apparently even computers agree with those judgments (or at least cluster "impressionists" in their own group - I didn't read the paper, but I expect that the cluster labels were added manually).

ETA: Got the paper. Excerpts:

The dataset includes 994 paintings representing 34 painters, such that each painter has at least 19 images in the dataset. The painters represent several different schools of art such as Early, High, and Northern Renaissance, Mannerism, Baroque, Rococo, Romanticism, Impressionism, Post and Neo Impressionism, Abstract Expressionism, Surrealism, and Fauvism, as commonly defined by art historians. The images were downloaded from various online sources, and normalized to a size of 640,000 pixels while preserving the original aspect ratio. The paintings that were selected for the experiment are assumed to be all in their original condition.

[...] To make the analysis more meaningful for comparing similarities between artistic styles of painters, we selected for each painter paintings that reflect the signature artistic style of that painter. For instance, in Wassily Kandinsky collection we included only paintings representing his abstract expressionism signature artistic style, and did not include his earlier work such as “The-Blue-Rider”, which embodies a different artistic style.

The dataset is used such that in each run 17 different paintings per artist are randomly selected to determine the Fisher discriminant scores of the features, and two images from each painter are used to determine the distances between the images using the WND method [Shamir 2008; Shamir et al. 2008, 2009, 2010]. The experiment is repeated automatically 100 times, and the arithmetic means of the distances across all runs are computed. [...]

The image analysis method is based on the WND-CHARM scheme [Shamir 2008; Shamir et al. 2008], which was originally developed for biomedical image analysis [Shamir et al. 2008, 2009]. The CHARM [Shamir, 2008; Shamir et al. 2010] set of numerical image content descriptors is a comprehensive set of 4027 features that reflect very many aspects of the visual content such as shapes (Euler number, Otsu binary object statistics), textures (Haralick, Tamura), edges (Prewitt gradient statistics), colors [Shamir 2006], statistical distribution of the pixel intensities (multiscale histograms, first four moments), fractal features [Wu et al. 1992], and polynomial decomposition of the image (Chebyshev statistics). These content descriptors are described more thoroughly in Shamir [2008] and Shamir et al. [2008, 2009, 2010]. This scheme of numerical image content descriptors was originally developed for complex morphological analysis of biomedical imaging, but was also found useful for the analysis of visual art [Shamir et al. 2010; Shamir 2012].

An important feature of the set of numerical image content descriptors is that the color descriptors are based on a first step of classifying each pixel into one of 10 color classes based on a fuzzy logic model that mimics the human intuition of colors [Shamir 2006]. This transformation to basic color classes ensures that further analysis of the color information is not sensitive to specific pigments that were not available to some of the classical painters in the dataset, or to the condition and restoration of some of the older paintings used in this study.

[...] As the figure shows, the classical artists are placed in the lower part of the phylogeny, while the modern artists are clustered in the upper part. A clear distinction between those groups at the center reflects the difference between classical realism and modern artistic styles that evolved during and after the 19th century.

Inside those two broad groups, it is noticeable that the computer was able to correctly cluster artists that belong in the same artistic movements, and placed these clusters on the graph in a fashion that is largely in agreement with the analysis of art historians. For instance, the bottom center cluster includes the High Renaissance artists Raphael, Da Vinci, and Michelangelo, indicating that the computer analysis could identify that these artists belong in the same school of art and have similar artistic styles [O’Mahony 2006].

The Early Renaissance artists Ghirlandaio, Francesca, and Botticelli are clustered together left to the High Renaissance painters, and the Northern Renaissance artists Bruegel, Van Eyck, and Durer are placed above the High Renaissance. Further to the right, close the High Renaissance, the algorithm placed three painters associated with the Mannerism movement, Veronese, Tintoretto, and El Greco, who were inspired by Renaissance artists such as Michelangelo [O’Mahony 2006]. Below the Mannerism painters the algorithm automatically grouped three Baroque artists, Vermeer, Rubens, and Rembrandt. Interestingly, Goya, Rococo, and Romanticism artist is placed between the Mannerism and the Baroque schools. The Romanticism artists, Gericault and Delacroix, who were inspired by Baroque painters such as Rubens [Gariff 2008], are clustered next to the Baroque group.

The upper part of the phylogeny features the modern artists. The Abstract Expressionists Kandinsky, Rothko, and Pollock are grouped together, as it has been shown that abstract paintings can be automatically differentiated from figural paintings with high accuracy [Shamir et al. 2010]. Surrealists Dali, Ernst, and de Chirico are also clustered by the computer analysis. An interesting observation is Fauvists Matisse and Derain are placed close to each other, between the Neo Impressionists and Abstract Expressionists clusters.

The neighboring clusters of Neo Impressionists Seurat and Signac and Post Impressionists Cezanne and Gauguin are also in agreement with the perception of art historians, as well as the cluster of Impressionists Renoir and Monet. These two artists are placed close to Vincent Van Gogh, who is associated with the Post Impressionism artistic movement. The separation of Van Gogh from the other Post Impressionist painters can be explained by the influence of Monet and Renoir on his artistic style [Walther and Metzger 2006], or by his unique painting style reflected by low-level image features that are similar to the style of Jackson Pollock [Shamir 2012], and could affect the automatic placement of Van Gogh on the phylogeny.