| <h2>Question</h2> | |
| <p><strong>To what extent can visual features alone predict the approximate historical period and cultural-geographic region of | |
| an artwork, and how well do these predictions generalise to artworks, artists and collections not encountered during | |
| training?</strong></p> | |
| <p>The question is experimental. That a model <em>produces</em> a prediction says nothing about whether the prediction is | |
| informative; Artifex is built to measure that, and to show the answer to users.</p> | |
| <h2>Scope</h2> | |
| <ul> | |
| <li><strong>Objects:</strong> paintings and painted imagery (including illustrated manuscript folios and album leaves), as catalogued by | |
| The Metropolitan Museum of Art and The Cleveland Museum of Art and released under CC0.</li> | |
| <li><strong>Targets:</strong> a broad cultural-geographic region (with narrower traditions where data allow) and an ordered date bin.</li> | |
| <li><strong>Input:</strong> the image only. Titles, artist names, inscriptions read by people, and other text are <em>not</em> used for | |
| prediction. Text search exists in the collection browser purely for browsing.</li> | |
| </ul> | |
| <h2>Hypotheses</h2> | |
| <p>H1 — <strong>Region is recoverable from appearance well above chance</strong> for traditions that are well represented in the | |
| reference collection, because format, materials, pictorial conventions and support (e.g. hanging scroll vs. canvas) | |
| differ systematically between traditions.</p> | |
| <p>H2 — <strong>Period is harder than region</strong>, and errors are mostly between <em>adjacent</em> bins: neighbouring centuries share | |
| conventions, and within-tradition change is gradual. Ordinal metrics (within-one-bin accuracy, error in years) should | |
| therefore be much better than exact-bin accuracy.</p> | |
| <p>H3 — <strong>Performance drops for unseen artists and more so for an unseen institution.</strong> Artist-disjoint splits remove style | |
| memorisation; collection-disjoint tests additionally change photography, colour processing and collecting history. | |
| The size of the drop indicates how much of the signal is collection-specific.</p> | |
| <p>H4 — <strong>Part of the signal is low-level</strong> (palette, varnish yellowing, paper tone, photographic conditions). A colour/texture | |
| baseline will beat the majority class, and greyscale input will reduce performance — but DINOv2 should outperform the | |
| low-level baseline if higher-level structure matters.</p> | |
| <p>H5 — <strong>Agreement between the classifier and nearest-neighbour retrieval is a useful reliability signal</strong>: predictions | |
| should be more accurate when the two agree.</p> | |
| <p>H6 — <strong>Calibration is necessary.</strong> Raw logistic-regression scores will be miscalibrated with balanced class weights; | |
| temperature scaling on validation data should reduce calibration error on test data drawn from the same collections, | |
| but cannot be relied upon for inputs from other distributions.</p> | |
| <h2>What would count as evidence</h2> | |
| <p>Each hypothesis is assessed against baselines run by the same code on the same splits (majority class, k-NN, random | |
| retrieval, ablation variants). The evaluation report of each run contains the numbers; this document deliberately | |
| contains none. Confirmations and refutations are recorded in the run's <code>report.md</code> and the Experiments page.</p> | |
| <h2>Relevance</h2> | |
| <p>Computational art history has studied style, influence and dating with learned representations (e.g. work on WikiArt | |
| style classification, the Replica project on visual pattern discovery, and large-scale analyses of museum collections). | |
| Artifex focuses on a narrower, practical setting — what a single photograph can tell a non-specialist about an unfamiliar | |
| painting — and on the honesty of that answer: abstention, calibration, provenance and leakage-aware evaluation.</p> |