Kurosawa Shot-Level Dataset: Shot Boundaries, Color, and Face-Derived Shot-Scale Features for 30 Films

Repository location

DOI: 10.5281/zenodo.18108847

Context

This dataset was produced as part of an independent research project on computational film-style analysis. As of the dataset release date (2025-12-31), no peer-reviewed publications based on this dataset have been published.

Steps

Figure 1 summarizes the automated pipeline for generating shot-level “syntax proxy” measures. We apply TransNet v2 (Souček & Lokoč, 2024) for shot boundary detection and discard shots shorter than 0.5 s. From each retained shot, five frames are sampled at fixed fractional positions (t = 1/6–5/6) and decoded with OpenCV (Bradski, 2000). We then extract (i) full-frame face statistics with RetinaFace (Deng et al., 2020), using the maximum face bounding-box height/frame height as a face-size proxy; and (ii) color metrics from a central ROI cropped by removing 12.5% from all four borders to reduce subtitles, watermarks, and border artifacts (Figure 2).

Figure 1

Figure 2

On the ROI, we compute HSV saturation (S) and CIELAB-derived brightness (L*) and warmth (b*–128) using a 1% trimmed mean, and average frame-level values to shots. On the full frame, RetinaFace yields per-frame face count and a face-size proxy; shot-level summaries include face coverage (frames with ≥1 face/5) and the median face-size proxy. Shot-scale labels are assigned only when face coverage ≥0.6 (≥3/5 frames); otherwise, shots are coded as No_Face/Other. We also export auxiliary variables (e.g., raw and truncated ASL, flags for shots >40 s, and log-transformed shot duration) with all shot records to CSV.

Sampling strategy

This dataset is a film-level census of all 30 feature films directed by Akira Kurosawa. For each film, TransNet v2 (Souček & Lokoč, 2024) produces a complete shot list with start/end timestamps and durations; shots shorter than 0.5 s are excluded from the released shot-level dataset and therefore from film-level summaries.

We report two film-level pacing statistics: raw ASL (mean duration of retained shots, ≥0.5 s) and a truncated ASL (mean after additionally excluding shots longer than 40 s) as a sensitivity analysis. Importantly, shots >40 s are kept in the shot-level dataset but flagged, as they may reflect genuine long takes or occasional boundary misses (e.g., gradual transitions). Reporting both values provides a conservative range and supports auditing of extreme-duration candidates.

For within-shot visual measurements, we use deterministic systematic sampling (Figure 3). From each retained shot, five frames are sampled at uniformly spaced relative positions (1/6–5/6) to avoid boundary-adjacent frames, and the same sampled frames are reused across feature types. Color features (HSV saturation; CIELAB-derived brightness and warmth) are computed per sampled frame and averaged to shot-level values.

Figure 3

From the same frames, RetinaFace (Deng et al., 2020) detects faces and records face count and the face-size proxy. A shot is labeled face-present if faces are detected in ≥3 of five frames (coverage ≥0.6). Shot scale is assigned from the median face-size proxy using thresholds: <0.13 (Long/Full), <0.25 (Medium), <0.45 (Medium Close-up), ≥0.45 (Close-up); otherwise it is coded as No_Face/Other.

To contextualize this proxy-based labeling choice, we contrast it with human-annotated alternatives, which differ in scope and scalability. CineScale (Savardi et al., 2021) provides frame-level labels across nine categories, and AVE (Argaw et al., 2022) annotates shot size as one of eight attributes using a professional workflow. ShotBench (Liu et al., 2025), meanwhile, benchmarks visual grammars for VLMs but still reports fine-grained inter-category confusion. Unlike these resources, our dataset prioritizes a scalable, filmography-wide operational proxy derived from face size. Intended for reproducible statistics rather than exhaustive taxonomy, we (i) gate labels by face coverage, (ii) retain a No_Face/Other code, and (iii) report both exact and adjacent agreement.

All analysis scripts and configuration files used to generate the dataset are publicly released to support reproducibility.

Quality control

As an external benchmark for pacing, we compared film-level ASL against Cinemetrics for all 30 films, prioritizing Barry Salt’s entries (n = 11) and otherwise using community-contributed values (n = 7); 12 films had no matching entry. To mitigate under-segmentation around gradual transitions, we report raw ASL (shots ≥0.5 s) and truncated ASL (excluding shots >40 s) as a plausible range. Of the 18 films with references, 16 fall within this range; the remaining discrepancies are flagged for manual audit (Table 1).

To assess internal reliability, we manually annotated 62 shots sampled across films and visual conditions (black-and-white vs. color; dialogue-driven vs. action). Shot-boundary detection achieved F1 = 0.91 within a ±0.5 s tolerance. Using the “face-present” rule (≥3/5 frames), face detection reached shot-level F1 = 0.96. Face-size–based shot-scale classification achieved 82% exact and 87% adjacent accuracy (±1 scale bin).

Taken together, the Cinemetrics comparison and our manual audit support the reliability of the released measures. Remaining discrepancies—often linked to gradual transitions or challenging face conditions—are handled via conservative reporting (raw vs. truncated ASL) and flagged for targeted review. Face-derived shot scale should be interpreted as an operational proxy.

Repository name

Zenodo

Object name

1. kurosawa_shot_level_data.csv;

2. Kurosawa30.zip;

3. generate_kurosawa_dataset.py

Format names and versions

CSV (UTF-8);

Python script (.py) [Python 3.10]

Creation dates

(2025-12-22) to (2025-12-29)

Dataset creators

Xueran Wu (Luoyang Normal University; principal investigator; data curation).

Language

English

License

Creative Commons Attribution 4.0 International

Publication date

(2025-12-31)

Notes

Kurosawa30.zip contains 30 per-film shot-level CSV files (one CSV per film).