Datasetpublic

szymonrucinski/types-of-film-shots

What a Shot! 2,919 film frames labeled with shot scale (8 classes), from film-grab.com. annotator count how labeled human 863 hand-labeled (gold) ai 2,056 DINOv2 classifier; low-confidence frames re-judged by Claude Opus (active learning) Columns image, label — ambiguous, closeUp, detail, extremeLongShot, fullShot, longShot, mediumCloseUp, mediumShot annotator — human or ai source — human / v2_highconf / opus_review / opus_resolved_amb… See the full description on the dataset page: https://huggingface.co/datasets/szymonrucinski/types-of-film-shots.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

9likes184downloads

Dataset Card

What a Shot!

2,919 film frames labeled with shot scale (8 classes), from film-grab.com.

`annotator`	count	how labeled
human	863	hand-labeled (gold)
ai	2,056	DINOv2 classifier; low-confidence frames re-judged by Claude Opus (active learning)

Columns

—image, label — ambiguous, closeUp, detail, extremeLongShot, fullShot, longShot, mediumCloseUp, mediumShot
—annotator — human or ai
—source — human / v2_highconf / opus_review / opus_resolved_amb
—confidence — labeler confidence (1.0 for human)
—movie — source film (AI rows)

Notes

—Confidence-banded review: the DINOv2 model labeled all frames; every frame with confidence < 0.8 was re-judged by Opus (it overruled the model on 53–65% of them — adjacent shot sizes are genuinely hard).
—Ambiguous cleanup: 62 of the original 66 human-ambiguous frames were resolved to a concrete shot scale on review (source=opus_resolved_amb); only 4 remain truly ambiguous.
—Use: filter(annotator=="human") for a clean eval set; use everything (optionally confidence-weighted) for training.