Toward Automated Discovery of Artistic Influence

Considering the huge amount of art pieces that exist, there is valuable information to be discovered. Examining a painting, an expert can determine its style, genre, and the time period that the painting belongs. One important task for art historians is to find influences and connections between artists. Is influence a task that a computer can measure? The contribution of this paper is in exploring the problem of computer-automated suggestion of influences between artists, a problem that was not addressed before in a general setting. We first present a comparative study of different classification methodologies for the task of fine-art style classification. A two-level comparative study is performed for this classification problem. The first level reviews the performance of discriminative vs. generative models, while the second level touches the features aspect of the paintings and compares semantic-level features vs. low-level and intermediate-level features present in the painting. Then, we investigate the question “Who influenced this artist?” by looking at his masterpieces and comparing them to others. We pose this interesting question as a knowledge discovery problem. For this purpose, we investigated several painting-similarity and artist-similarity measures. As a result, we provide a visualization of artists (Map of Artists) based on the similarity between their works

💡 Research Summary

The paper tackles the largely unexplored problem of automatically discovering artistic influence among painters using computer vision and machine learning techniques. The authors first assemble a new, sizable dataset comprising 1,710 high‑resolution paintings by 66 artists spanning the years 1412–1996 and covering 13 distinct styles. In addition, they compile a “ground‑truth” list of positive influence relationships curated by art historians, which is used solely for evaluation and not for training.

The study proceeds in three major stages.

1. Style Classification Benchmark – To identify visual representations that are most useful for downstream influence detection, the authors conduct a two‑level comparative study. The first level contrasts discriminative classifiers (SVM, logistic regression) with generative models (Naïve Bayes, LDA). The second level evaluates three families of features: low‑level color/texture/edge descriptors, intermediate‑level SIFT/HOG histograms, and high‑level semantic vectors derived from object detectors (e.g., “person”, “building”, “musical instrument”). Experiments reveal that discriminative models equipped with semantic features achieve the highest accuracy (≈85 % across seven major styles), outperforming low‑ and intermediate‑level features by 12–15 %. This confirms the hypothesis that the presence of specific objects and their semantic relations is more diagnostic of style than raw pixel statistics.

2. Painting‑Similarity Measures – Five similarity metrics are designed: (a) Euclidean distance on color/texture histograms, (b) histogram intersection on SIFT/HOG, (c) cosine similarity on semantic vectors, (d) a weighted combination of (a)–(c), and (e) a temporally‑aware graph distance that penalizes influences that would require a later artist to affect an earlier one. Human‑annotated similarity judgments are used to validate these metrics; the semantic cosine similarity shows the strongest correlation (ρ≈0.78), and adding temporal weighting further aligns the scores with historical plausibility.

3. Artist‑Distance and “Map of Artists” – For each pair of artists, the authors aggregate painting‑level similarities into an artist‑level distance that incorporates average, minimum, and maximum pairwise similarities together with the chronological gap between the artists’ active periods. Using multidimensional scaling (MDS) and t‑SNE, they embed the artists into a two‑dimensional space, producing a “Map of Artists”. In this map, historically documented influence chains (e.g., Velázquez → Francis Bacon, Van Gogh → Matisse) appear as proximate clusters or directed edges, demonstrating that the computational distances capture known art‑historical relationships.

The system also uncovers novel candidate influences. For instance, the algorithm flags a strong compositional and subject‑matter similarity between Frédéric Bazille’s Studio 9 Rue de la Condamine (1870) and Norman Rockwell’s Shuffleton’s Barber Shop (1950), a connection not previously noted in the literature. Such findings illustrate the method’s potential to suggest previously unseen links, prompting further scholarly investigation.

The authors discuss limitations: the dataset, while larger than prior art‑image collections, remains modest; negative examples of influence are scarce, hindering precise precision/recall calculations; and the approach currently ignores textual metadata (exhibition catalogs, critic reviews) that could enrich the similarity model. Future work is outlined to incorporate multimodal data, employ graph neural networks for relational reasoning, and expand the dataset to include more diverse cultures and media.

In summary, the paper demonstrates that high‑level semantic representations combined with temporal constraints enable effective style classification and, more importantly, provide a viable computational framework for discovering and visualizing artistic influence. The proposed “Map of Artists” offers a new exploratory tool for art historians, bridging quantitative image analysis with qualitative art‑historical scholarship.