Can a model recognise visual motifs in images, and what does the published work actually show?

Phillips et al. (UPF, arXiv v1, 21 Oct 2024) train a shallow classifier on CLIP features to tag images with 20 expert-chosen motifs (e.g. Mirror, Stairs, Woman in the Window) and report F1 0.9138 on their own test set, with 88% of test images predicted exactly. The dataset (10,760 images) was not public at v1 and the paper does not test whether a motif carries a theme in a work. It is a paper result on image-level composition recognition, not a released tool. Curator qualifications: Curator located the evidence excerpts in the work copies (automated prefix match; the one apparent miss, S25-05 E4, is a curly-apostrophe variant and is present). Own-test-set results with expert-chosen motifs; dataset not public at v1. AI connection: documented research application, not validated on film footage. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/html/2410.15866v1",
      "note": "results; data"
    }
  ]
}