Glossary
Terms appear in the order you are likely to meet them. Field names in
backticks refer to keys in data/*.jsonl.
The task
Hero frame — a single still image taken from within a movie that would work as its cover, thumbnail, or promotional entry point. This exact sentence is the definition given to every judge and reader, and it is the sentence handed to multimodal methods as their query.
Key-frame selection — the task of producing such a still from a film. This benchmark evaluates it; it does not perform it.
The unit of measurement
Chain — six stills from one film, frozen in a definite order from best hero
frame to worst. A chain is what a candidate frame gets inserted into, and the
whole benchmark is 2,031 of them. Field: each entry of chains[].
Anchor, also positive — the one image in a chain that a human chose: an
IMDb publicity still (type == still_frame). It is a prior, not ground truth.
It participates in the ordering on equal terms and can lose. Field:
chains[].anchor, and order[].is_anchor.
Sampled frame, also negative — one of the other five images, taken from
the film's own LSMDC clips by dividing each clip into 2.5-second buckets and
sampling the midpoint. Field: order[].from_clip, order[].clip_t_sec.
Group, also draw — one anchor plus five sampled frames, before ordering. Each anchor had up to four mutually disjoint groups prepared; the one that produced the most decisive comparisons became the chain.
Coverage — how much of the film's visual variety a group spans, computed as
the mean over the negative pool of each frame's cosine similarity to its
closest match among the group's five. Higher is broader. Field:
chains[].coverage.
How images were ordered
Pairwise BT — every one of the C(6,2) = 15 pairs in a group is judged twice, once in each left-right placement, and Bradley-Terry is fitted to the outcomes to produce a strength per image.
Gamma (γ) — the fitted Bradley-Terry strength. Larger means more
hero-frame-like. Field: order[].gamma.
Theta (θ) — γ mapped linearly onto 0–100 for display only. Do not treat
differences in θ as calibrated. Field: order[].theta_0_100.
Decisive, also confident — the judge picked the same image under both
placements. Field: pairs[].status == "confident".
Unresolved, also flip — the judge picked different images under the two
placements, contradicting itself. Counted as an error everywhere, with no
partial credit. Field: pairs[].status == "unresolved".
Confident rate — the fraction of a chain's 15 pairs that are decisive.
Corpus mean 0.953; 52% of chains are decisive on all 15. Field:
chains[].stability.confident_rate.
The rubric
Rubric — the set of comparison guidelines the judge read while ordering one
film's images. Every rubric was learned for its own film, starting from an
empty set. No human wrote any of them. Field: rules[] in rubrics.jsonl.
Criterion — one guideline. Rendered to the judge in a fixed shape:
Only when both images {applicability}: prefer {preferred_state} over {dispreferred_state}.
Field: rules[].as_the_judge_read_it is the sentence verbatim.
Applicability — the condition both images must meet for a criterion to
apply to their comparison. If either falls outside it, the criterion is skipped
for that pair, and neither image is penalised for falling outside. Mandatory:
a criterion without one is invalid. Field: rules[].applicability.
Provenance — which image pair surfaced a criterion, which image the judge
picked as better, and which group the criterion was cross-checked against.
Field: rules[].provenance.
The two admission gates
Both had to pass before a proposed criterion entered a rubric. In the second round they rejected 94.4% of proposals.
Inverted attack (gate i) — insert the candidate criterion, re-rank the group it came from, then rewrite the criterion with its two preference ends swapped (the applicability clause is left byte-identical) and re-rank again. A criterion that genuinely governs the comparison must be able to flip it.
s_stab / s_flip — the two scores from that gate. 1.0 means the intended image became the decisive winner, 0.5 means it merely led on γ, 0 means it lost. Admission required both above zero and at least one at 1.0.
Cross-group check (gate ii) — insert the candidate, then re-judge the already-decisive pairs of the film's most dissimilar group. Disturbing too many of them rejects the criterion.
xg_flips — how many previously decisive pairs the candidate disturbed in that check.
Freezing
Phase B — the final pass. Re-rank every group under the finished rubric, delete any anchor that lands in the bottom half, backfill a replacement, prune the criteria that came from deleted anchors, and freeze what remains into chains.
majority_reversal — the reason recorded when an anchor was deleted: its
rank came out at 4 or worse out of 6. Happened 451 times. Field:
removed_anchors[].reason.
pruned_by_phase_b — criteria removed because their source anchor was deleted. They are recorded but are not part of the final rubric. 70 across the corpus.
Scoring
Insertion — the scoring operation. The reader sees the six chain images already in order plus one candidate, and answers what position the candidate takes. Asked twice, with both the image order and the description reversed the second time, then averaged.
s — the score of one candidate against one chain: the fraction of the six chain images the candidate was placed above. Range 0 to 1.
Soft insertion rank — the same quantity as a position: 7 - 6s. One
position equals 0.167 of s.
S — a method's score, averaged over chains, then over the method's output frames, then over films. Random is 0.501, human publicity stills 0.722.
S@k — S recomputed using only the method's first k output frames, in the method's own ordering. Measures how well a method ranks its own output.
top-1 rate — how often a candidate was placed above all six chain images.
beat-the-still rate — how often a candidate was placed above the chain's human publicity still specifically.
Journey percentage — where a method sits on the scale from random (0%) to human publicity stills (100%). Negative means below random.
Reproducing the candidate pool
Canonical frame track — the single 1 fps frame pool every method selects from, so the comparison is about selection rather than decoding density. About 500,000 frames across the corpus, shipped as coordinates rather than pixels.
Concatenated timeline — clips sorted by (start_sec, clip_id) and laid end
to end. t_cat is a position on it; t_movie is the position in the original
film.
Bucket midpoint — the sampling rule. A clip of duration d yields
n = max(1, floor(d)) frames, the i-th at (i + 0.5) * d / n. Every clip
contributes at least one frame, including clips under a second.
dHash — the 64-bit perceptual hash used to confirm a rebuilt pool matches. Stable across the resolution difference between our stored frames and a fresh native-resolution decode, so a Hamming distance at or below 8 counts as a match.