# Feature extractors

One script per family. All read `data/artwork_metadata.csv` + `data/images/`, write to
`data/features/`, checkpoint regularly, and resume if re-run.

| family       | script                    | dims | what it measures |
|--------------|---------------------------|-----:|------------------|
| hand-crafted | `extract_handcrafted.py`  |  380 | color (HSV joint + grey-world + per-channel histograms), light (brightness/contrast/darkness/edge density), symmetry, flatness (SLIC superpixel LAB uniformity), texture (FFT radial bands + 59-bin nri-uniform LBP), lines (Hough count/density, straight ratio, edge-angle histogram) |
| CLIP         | `extract_clip.py`         | 512 + 27 | ViT-B/16 image embedding + zero-shot cosine scores against 27 hand-written religious-attribute prompts |
| DINOv2       | `extract_dino.py`         |  768 | self-supervised CLS token — visual style/appearance similarity |
| faces        | `extract_faces.py`        |   39 | YuNet detections summarized: count, coverage, sizes, positions, tilt/frontality, arrangement, spatial histograms |
| pose         | `extract_pose.py`         |   39 | YOLOv8-pose, **largest figure only**: torso-normalised 17-keypoint skeleton + figure area + detection-quality columns + person count |

Notes:

- Hand-crafted features use the corrected preprocessing pipeline: images are
  border-cropped, guardedly background-masked (see `preprocessing/`), and
  capped at 1024 px on the longest side so texture/edge features measure the
  artwork rather than the museum's scan resolution.
- The pose vector stores quality indicators (`main_kpt_conf`, `main_torso_px`)
  because photo-trained detectors are unreliable on stylised bodies — filter
  on them before interpreting posture.