weitaikang/Weitaikang_bench_data

HeroFrame-Bench

A benchmark for key-frame selection that scores a selected frame directly, without routing it through a question-answering model and without requiring it to match a fixed reference set.

204 films, 2,031 frozen comparison chains, 1,970 learned criteria, 30,465 pairwise judgements.

---

What problem this addresses

A film is almost always encountered first as a single still: a cover, a thumbnail, a poster. Producing that still from the film is the task called key-frame selection. This benchmark calls the target of that task a hero frame — a single still image taken from within a movie that would work as its cover, thumbnail, or promotional entry point.

Evaluating key-frame selection is where the difficulty sits, and the two existing approaches each lose the thing being measured.

Delegating to question answering. Feed the selected frames to a frozen vision-language model, ask it questions about the video, report accuracy. The resulting number is a joint property of two systems: a weak reader punishes a good selection, a strong reader forgives a redundant one. Worse, the questions have to be machine-checkable, so they can only probe retrievable content — precisely not the part of hero-frame quality that is in doubt.

Comparing against a fixed answer. Score against a reference summary, or against photographic proxies like brightness and sharpness. A good frame outside the reference set is recorded as a miss; and exposure and focus are not what makes an image work as a cover.

Both miss the trade-off that hero-frame selection actually involves: the most informative frame is often visually unappealing, and the most eye-catching frame is often a minor character's small moment.

How this benchmark scores instead

Each film comes with roughly ten chains. A chain is six stills from that film, frozen in a definite order from best hero frame to worst. One of the six is a human-chosen publicity still; the other five are frames sampled from the film itself.

To score a candidate frame, the frame is inserted into a chain and the reader is asked what position it takes. The score is the fraction of the chain the candidate was placed above:

s(F, C) = (number of chain images F was placed above) / 6      in [0, 1]

Nothing has to match a reference. A frame that appears in no reference set can still earn a high score on its own merits, and the score never passes through a downstream QA model.

What the numbers look like. Random frames land at 0.501, against the 0.5 implied by construction. Human publicity stills reach 0.722. The strongest method measured so far reaches 0.586 — 38% of the way from random to human.

Where the ordering came from

The chains are not ordered by human annotators, and not by an off-the-shelf notion of image quality. Each film's ordering was produced by a judge model reading a rubric learned for that film — a set of comparison guidelines, starting from nothing, written by the model itself while it was ordering the images.

Every learned guideline had to survive two falsifiable admission gates before entering the rubric: written in reverse it must actually flip the comparison it came from, and adding it must not disturb already-settled comparisons in the film's most dissimilar group. In the second round these gates rejected 94.4% of proposals.

Whether that machinery works is an empirical question, and the answers are in the paper. Two of them matter for trusting this data: against 450 pairs judged blind by humans, the rubric-guided judge agrees 83.78% of the time versus 77.56% with no rubric; and serving a film another film's rubric is statistically indistinguishable from serving no rubric at all (p = 0.41), so the gain is not generic photographic common sense.

---

Getting started

Evaluation needs film footage, which cannot be redistributed. The steps below get you from nothing to a comparable score.

1. Install the evaluation code

hf download weitaikang/Weitaikang_bench_github --repo-type dataset --local-dir HeroFrame-Bench
cd HeroFrame-Bench && pip install -e ".[all]"

Weitaikang_bench_github holds the reference implementation of every step here, plus all eleven baselines. The third repository, Weitaikang_bench_paper, holds the evidence behind the paper — you do not need it to use the benchmark.

2. Download this dataset

hf download weitaikang/Weitaikang_bench_data --repo-type dataset --local-dir ./heroframe-bench

About 2.8 GB. After this you have the chains, their images, the rubrics, and the frame-track index. You can already read the benchmark and inspect what it considers a good frame.

3. Obtain the film footage

The films come from LSMDC (Large Scale Movie Description Challenge). Request access at https://sites.google.com/site/describingmovies/. This is the only piece you have to source yourself, and it is unavoidable: the footage is licensed.

4. Rebuild the candidate frame pool

Every method draws frames from one shared pool, so that the comparison is about selection strategy rather than about whose decoder samples on a denser grid. The pool is ~500,000 frames at 1 fps; frame_track/track_index.jsonl.zst pins each one to a (clip_id, t_in_clip) coordinate.

python -m heroframe.scripts.rebuild_track --lsmdc /path/to/lsmdc --out ./track
python -m heroframe.scripts.verify_track  --track ./track

verify_track recomputes a perceptual hash on 24 frames per film and compares against frame_track/verify_dhash.jsonl.zst. It catches the failures that matter — clips concatenated in the wrong order, a different bucket midpoint, a silently skipped clip — while tolerating harmless encoder differences. Details in frame_track/PROTOCOL.md.

Budget several hours for a full rebuild; it resumes if interrupted. Start with --movies 0001_American_Beauty 0003_CASABLANCA 0008_Fargo to confirm the path works before committing to all 204.

5. Run your method

Your method receives only the film's footage. No rubric, no reference stills, no metadata. Multimodal methods additionally receive one fixed sentence, identical for every film, in eval/query.txt.

Output at most 5 frames per film, as frame indices into the track:

{"movie_id": "0001_American_Beauty", "frames": [{"rank": 1, "frame_idx": 2477}, ...]}

You choose how many frames to submit. Scoring averages rather than takes the best, so a fifth mediocre frame can pull your score down.

6. Score

python -m heroframe.scripts.evaluate --selections my_method.jsonl --track ./track

Reports S, S@k, top-1 rate, and beat-the-still rate, alongside the eleven official rows. Use the same reader the benchmark specifies (Qwen3.6-27B); frame-level scores are not comparable across readers, which is a measured property rather than a caution — see the reliability analysis in the paper.

---

What is in this repository

data/ — one line per film

File	Lines	Contents
chains.jsonl	204	2,031 chains. Each carries the six-image order with Bradley-Terry strengths, all 15 pairwise judgements with their status, and stability statistics.
rubrics.jsonl	204	1,970 criteria. as_the_judge_read_it is the exact sentence the judge saw. Each records its source image pair and which round learned it.
movies.jsonl	204	Title, year, synopsis, characters, key events, TMDB record.
clips.jsonl	204	128,085 LSMDC clips: timecodes, duration, caption, and which sampled frames each contributed.

images/<movie_id>/<image_id>.jpg

11,161 images at native resolution — every image appearing in a frozen chain. Publicity stills carry rm-prefixed ids; sampled frames are named for their source clip and timecode. images/image_sha256.jsonl lists a hash per file.

frame_track/

The 1 fps candidate pool as coordinates rather than pixels, plus the verification hashes. See PROTOCOL.md.

eval/

judge_prompt_example.md is the scoring prompt rendered verbatim on a real film. query.txt is the sentence handed to multimodal methods. baselines.jsonl holds the eleven official rows: aggregate scores, per-film scores, and the actual frames each method picked.

stats/corpus_stats.json

Genre distribution, year and runtime spread, rubric size distribution, chain statistics, and the image-property correlations underpinning the analysis.

---

Reading the data correctly

The publicity still is a prior, not ground truth. It participates in the ordering on equal terms. A sampled frame ranking above it is preserved as real signal — that happens in 48% of chains. The still's only privilege is a floor: if it falls into the bottom half it is deleted and replaced, which happened 451 times across construction.

pairs[].status distinguishes two things. confident means the judge chose the same image under both placements. unresolved means it contradicted itself; those are counted as errors throughout, with no partial credit. 95.3% of the 30,465 pairs are confident.

theta_0_100 is for display. It is the Bradley-Terry strength linearly mapped to 0–100. gamma is the raw strength.

Rules pruned in Phase B are not in the rubric. They appear under pruned_by_phase_b because their source anchor was deleted.

Position in the film is not a confound. Sampled frames average 0.545 in normalised film time and stay flat across chain rank, so rank is not a proxy for when something happens.

A glossary of every term is in GLOSSARY.md.

---

Limitations worth knowing before you use this

• The ordering comes from a model, not from humans. Human agreement was measured on 450 pairs, not on the full corpus.
• Each human pair carries one vote, so inter-annotator agreement cannot be computed. The 83.78% figure is agreement with a single person's choice.
• The positive end is IMDb publicity stills, themselves products of a marketing process with its own conventions.
• 204 English-language films from LSMDC, 1943–2014. Broad genre coverage, but the distribution is Western theatrical cinema. No episodic content, no non-Western cinema, no user-generated video.
• The reader model is part of the specification. Results are reproducible and comparable with the specified reader; frame-level scores across readers are not.

---

Licence and image rights

Images are redistributed for non-commercial academic research only. Publicity stills originate from IMDb; sampled frames derive from LSMDC clips. See LICENSE.md. Actor portraits used during construction come from the MovieBench Character Bank and are not included; they are not needed for evaluation.

Citation

@inproceedings{heroframebench,
  title     = {HeroFrame-Bench: Evaluating Key-Frame Selection Without
               Delegating to Question Answering},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2027}
}