Weitaikang_bench_github / LEADERBOARD.md
weitaikang's picture
HeroFrame-Bench evaluation code
425512c verified
|
Raw
History Blame Contribute Delete
4.18 kB

Leaderboard

All rows: reader Qwen3.6-27B, insertion form D, all chains per film, at most 5 frames per film, 204 films. 110,685 (frame, chain) units in total.

# Method S 95% CI Insert rank S@1 Top-1 Beats the still Journey
Human IMDb still ceiling 0.7218 [0.711, 0.733] 2.67 0.735 13.4% 35.1% 100%
1 CueKFS j7 (gpt-5.5, budget 96) tuned 0.5855 [0.566, 0.605] 3.49 0.588 6.5% 23.6% 38.2%
2 CueKFS j7 (gpt-5.5, budget 128) tuned 0.5795 [0.560, 0.599] 3.52 0.568 6.4% 23.1% 35.4%
3 CueKFS j7 (gpt-5.5, budget 64) tuned 0.5785 [0.560, 0.597] 3.53 0.585 6.2% 22.8% 35.0%
4 CueKFS j7 (gpt-5.5, published) 0.5653 [0.547, 0.584] 3.61 0.572 5.7% 21.6% 29.0%
Largest-face heuristic probe 0.5571 [0.537, 0.578] 3.66 0.500 5.0% 20.2% 25.3%
Random floor 0.5015 [0.486, 0.517] 3.99 0.489 2.9% 16.1% 0%
5 Uniform sampling 0.4385 [0.422, 0.455] 4.37 0.330 2.6% 13.6% −28.6%
6 SigLIP2 text retrieval 0.3620 [0.341, 0.383] 4.83 0.341 1.1% 8.3% −63.3%
Brightness + sharpness probe 0.2969 [0.274, 0.320] 5.21 0.266 1.3% 6.5% −92.9%
7 WFS-SB (CVPR'26) 0.2633 [0.244, 0.283] 5.42 0.248 1.0% 5.4% −108.1%

Confidence intervals treat each film's S as one observation (n = 204). ceiling and floor are reference points, not methods. probe marks single-attribute selectors included to show what the score is made of.

Against the random baseline

Paired Wilcoxon signed-rank on the 204 per-film scores:

Method ΔS p
Human still +0.220 9.0e−34 above
CueKFS j7 (budget 96) +0.084 1.8e−12 above
CueKFS j7 (published) +0.064 5.9e−08 above
Largest-face heuristic +0.056 8.9e−06 above
Uniform sampling −0.063 6.7e−07 below
SigLIP2 text retrieval −0.139 2.3e−19 below
Brightness + sharpness −0.205 2.8e−26 below
WFS-SB −0.238 1.0e−31 below

Reading the table

Only one non-trivial method beats random. And its advantage over a plain face detector is not measurable in the published configuration (ΔS = +0.008, p = 0.63) — only the swept configuration separates (p = 0.032).

Three methods score significantly below random. They are not broken; they optimise something else. Uniform maximises temporal coverage, SigLIP2 maximises text-image similarity, WFS-SB detects semantic boundaries and spreads budget across segments. When those objectives are orthogonal to hero-frame quality, "below random" is the informative answer, and the benchmark says it with a sign rather than a shrug.

Nobody covers 40% of the distance. The gap that remains is 0.136 of S, or 0.818 of an insertion position. For scale, sweeping CueKFS's single most important hyperparameter is worth 0.020 — the headroom is not in tuning.

The scale behaves. Random landing at 0.501 against a construction-implied 0.5 means the zero point was not fitted.

What the remaining distance is not

Photographic quality. Across 10,200 selected frames, the correlation between a frame's score and its brightness, contrast, sharpness, colourfulness, or edge density is |ρ| ≤ 0.09. A selector optimising exposure and focus (bright_sharp) scores below random.

Face size is the one attribute that correlates strongly (ρ = 0.485), but it is a threshold rather than a dial, and it has a ceiling: human publicity stills carry a mean max-face fraction of 0.045, one eighth of what the face heuristic selects for, and score 0.165 higher.

Submitting

Open a pull request adding your row, with the result JSON from scripts/04_evaluate.py. See examples/submit_your_method.md.

Requirements: the specified reader, the unmodified prompt, a verified frame track, and — if your method calls an LLM internally — the model named in the method name and reported in its published configuration.