PaperlayerView exact revision on arXiv

Source-linked AI summary

Measuring the evolution of camera distance across a century of film

David Bamman, Allison Cooper, Dan Hickey, Madison Mar

arXiv:2609.31734v1cs.CV

TL;DR

The paper examines how camera distance changes over film history and how gender and animation relate to shot distance. It measures apparent distance across a 5,205-film corpus, finding shifts around sound and television, gender disparities in live-action close-ups, and no comparable gender effect in animation.

  • Problem

    The paper examines how camera distance changes over time, who is shown in medium close-ups, and whether animated films use distance differently.

    Source: §3 · p. 6, §3.2 · p. 7, §3.3 · p. 9

  • Method

    The study operationalizes apparent distance into shot categories and analyzes a corpus of 5,205 films viewed in theaters from 1922–2025.

    Source: §1.3 · p. 3, §2 · p. 5

  • Results

    Camera distance shifted around sound and television but had no statistically relevant smartphone effect; women receive more medium close-ups in live action, while animation shows no such gender effect.

    Source: §4 · p. 10

  • Takeaways & Limitations

    The findings suggest that camera-distance grammar may be more resilient across platforms than previously thought, while gender asymmetry extends across live-action cinema through the present.

    Source: §4 · p. 10

    Apparent framing does not establish physical camera distance, and subject ambiguity or nonstandard lenses can complicate its interpretation.

    Source: §2 · p. 4

Abstract

from arXiv · show

The rise of computer vision and artificial intelligence has made possible new forms of large-scale computational measurement. We apply these techniques to a deep collection of 5,205 digitized films viewed in theaters between 1922-2025 (covering popular, prestigious, and independent movies) to trace the development of one of the most fundamental ways through which film communicates: by manipulating the space between the camera and its subject. This work finds abrupt changes with the rise of new technologies in sound and television, and allows us to shed an empirical light on gender disparity (women, despite having substantially less screentime than men, are disproportionately the subject of closer shots), and illustrate how animated films both inherit and break free from the norms of live-action filmmaking.

1 Data

The study assembles popular, prestige, and independent films into a 5,205-film corpus spanning 1922–2025, then operationalizes apparent camera distance for corpus-scale measurement. Its data sources, annotation, model evaluation, and validation define the scope and basis of those measurements.

1.1 Popular

  • Popular films are selected from yearly box-office rankings, up to 50 per year, yielding 4,064 unique movies.The sources are Box Office Mojo for 1980–2025 and Variety-derived data for 1922–1979.Source: §1.1 · p. 3

1.2 Prestige

  • The prestige collection contains 1,883 unique films ranked using nominations from 12 award organizations and selections from prestigious film rankings.Acquisition prioritized films with the most nominations among the organizations.Source: §1.2 · p. 3

1.3 American Independent

  • The independent collection consists of 226 American films listed in Directory of World Cinema: American Independent.This sample offers a perspective not directly aligned with popularity or prestige.Source: §1.3 · p. 3

1 Data

  • 5,205 unique films viewed in theaters from 1922–2025 combine popular, prestige, and independent collections; not all films could be acquired, and availability reflects later popular and commercial interests.The sources offer different vantage points on camera-distance use.Source: §1.3 · p. 3, §1 · p. 3

2 Measuring distance

  • The study measures apparent distance—how far away a subject appears—rather than unknowable physical camera distance, using seven ordinal categories and a separate intertitle category.Framing can differ from physical distance, and wide-angle or telephoto lenses complicate its interpretation.Source: §2 · pp. 4–5
  • Annotators labeled distance categories across 4,939 shots sampled from 312 films spanning 1922–2025 and 100 additional animated films.Up to one minute per film was sampled; moving shots were labeled at the start and end of each applicable distance category.Source: §2 · p. 5
  • Base SigLIP2 achieved 68.9% accuracy and 95.61% within-one accuracy; prevalence estimates were corrected using prediction-powered inference in centered 25-year windows.The reported accuracies are ±3.3 and ±1.0, respectively; corrected trends broadly remained the same with wider confidence sets.Source: §2 · p. 5
  • Across 89 films, predicted distance proportions agreed with Barry Salt’s annotations at an average Spearman rank correlation of ρ = 0.880.The comparison maps Salt’s taxonomy onto seven ordinal categories.Source: §2 · p. 5

3 Findings

Across popular films, shot distance shortened around the introductions of sound and television, and these broad temporal patterns also held across collections, genres, and aspect ratios. Women received more medium close-ups than men in live-action films, while animation used more long shots and showed no significant gender difference in closer shots.

3.1 Camera distance over time

  • +0.054 in M and −0.081 in long shots followed sound’s introduction; television’s rise brought +0.044 in MCU and −0.060 in ML, while smartphone viewing showed no significant change.Across the period, medium close-ups more than doubled during television’s near-widespread adoption, while medium shots remained largely stable.Source: §3.1 · p. 7
  • Broad temporal patterns were largely identical across popular, prestige, and indie collections and held within the major genres and aspect ratios examined.Indie movies may contain slightly closer shots than contemporaneous popular or prestige movies.Source: §3.1 · p. 7

3.2 The male gaze: Women are seen through more medium close-ups than men

  • Women had 20% (+5 percentage points) more medium close-ups than men despite significantly less screentime, and the disparity persisted throughout the study period.The difference appeared across popular, prestige, and indie collections, with somewhat stronger differences in earlier cinema.Source: §3.2 · p. 8, Figure 3 · p. 8

3.3 Animation shows longer distance, and no gendered effects

  • Animation contained over twice as many long shots as live action, while live-action films concentrated on medium close-ups and medium shots.The long-shot pattern partly reflects animation’s freedom from camera-positioning constraints and can situate small non-human characters within larger environments.Source: §3.3 · p. 9, Figure 4 · p. 9
  • In animation, no statistically significant closer-shot gender differences appeared for either human or non-human characters; the medium-close-up disparity was specific to live action.The paired comparison included live-action characters, human animated characters, and non-human animated characters.Source: §3.3 · p. 9

4 Discussion

Camera distance changed with the introduction of sound and television, but not statistically following smartphones; gendered differences persist across live-action filmmaking, while animation differs. The findings also show how computational analysis can complement close viewing while risking simplification of cinematic complexity.

4 Discussion

  • Sound and television coincide with confirmed changes in camera distance, while smartphones show no statistically relevant effect.Widescreen films responding to television had the greatest proportion of extreme long shots, often favoring landscape and travel imagery.Source: §4 · p. 10
  • The lack of a smartphone effect suggests camera-distance grammar may be more resilient across platforms than previously thought.The authors also raise the possibilities that closer-view trends have played out or that digital change affects viewing environments more than formal properties.Source: §4 · p. 10
  • Women receive far more medium close-ups than men, and this gender asymmetry extends beyond classical Hollywood into independent live-action films.The strong live-action effect is not seen in animation and may be rooted partly in the objectifying function of closer shots.Source: §4 · p. 10
  • Computational analysis can complement close viewing and extend measurement to other visual forms, but formal operationalization risks simplifying filmmakers’ creative resistance to convention.Potential extensions include camera movement, angle, and blocking.Source: §4 · pp. 10–11

5 Materials and Methods

The study combines a broad film collection, supervised shot-distance classification, actor recognition, and gender metadata to measure camera distance and its associations across films.

5.1 Data

  • The dataset contains popular, prestige, and independent films from 1922 onward, with earlier estimates less certain because fewer films survive or are available.The popular collection is limited to theaters and cities reported by Variety, and many silent films from the 1920s are lost.Source: §5.1 · p. 11

5.2 Shot classification

  • Shot distance is treated as supervised classification against human judgments, using annotations from one-minute samples of 412 films, including 100 animated films.Annotators labeled distances per shot; disagreements were adjudicated, and clips with multiple distances received time-span labels.Source: §5.2 · p. 11
  • 68.9% accuracy and 95.6% within-one accuracy were achieved by SigLIP 2 base 384px, with errors generally at category boundaries.The model was selected from several architectures using a development set.Source: §5.2 · p. 11

5.3 Actor recognition

  • Actor recognition matches detected on-screen faces to cast images, using separate live-action and animation pipelines.The live-action pipeline uses YuNet, face tracks, and InsightFace representations; the animation pipeline uses RT-DETR and DINOv2.Source: §5.3 · p. 12

5.4 Gender

  • Live-action gender labels reflect Wikidata perceptions of actors near a film’s release, while animated-character labels draw on voice-actor gender and fan references.These labels describe actors rather than necessarily the characters they portray in live-action films.Source: §5.4 · p. 12

Funding

The research was supported by the Humanities and AI Virtual Institute.

Funding

  • The Humanities and AI Virtual Institute (HAVI), a Schmidt Sciences program, supported this research.Source: Funding · p. 12

Supplementary Information

The annotation guide classifies apparent camera-subject distance into categories and adjusts judgments to the scale depicted. It also specifies how to label movement and clips containing multiple shots.

S1 Distance taxonomy

  • The guide divides apparent camera-subject distance into eight categories, from extreme close-up through extreme long shot, plus intertitles.The listed distance bands range from under a foot for extreme close-ups to over 25 feet for extreme long shots.Source: S1 Distance taxonomy · p. 14
  • When a moving shot sustains multiple distances for about a second or more, annotators select each applicable category.Source: S1 Distance taxonomy · p. 14
  • If a clip contains multiple shots, annotators label all distances visible across those shots.Source: S1 Distance taxonomy · p. 14
  • For miniature or nonhuman-scale worlds, annotators judge distance relative to the depicted scale rather than ordinary human scale.A head-to-toe view of a miniaturized person can count as a long shot, not an extreme close-up.Source: S1 Distance taxonomy · pp. 14–15

S2 Total features over time

The number of feature films increased roughly linearly through 2000, then grew exponentially in the new millennium.

S2 Total features over time

S3 Size of digitized collection

The digitized collection spans 1922–2025 across popular, prestige, and American independent films, with one film from 1916. The supplementary figures document collection size over time and contextualize sound and television adoption.

S3 Size of digitized collection

S4 Technological change

S5 Model comparison

The supplementary material describes model-comparison and genre tables, then examines distance prevalence by genre and aspect ratio. Genre and aspect-ratio patterns do not strongly differ from overall prevalence, although some categories shift within particular ratios over time.

S5 Model comparison

S6 Genre

  • No genre displays a strongly different pattern from overall distance-category prevalence.Figure S7 plots the four major distance categories by genre; Table S2 reports medium-close-up shares by genre with 95% bootstrap confidence intervals.Source: S6 Genre · p. 18, Table S2 · p. 18, Figure S7 · p. 19
  • No aspect ratio displays a strongly different pattern from overall prevalence, though medium close-ups rise and medium long shots decline from 1950–1975.Long-shot rates remain relatively stable after an initial decline around 1930; aspect ratios are generally dominant in only one period.Source: S6 Genre · pp. 18–19, Figure S9 · p. 20

S7 Prevalence

Figures S10–S13 plot distance-category prevalence over time for popular and prestige movies, with common-scale and detailed views. All include 95% bootstrap confidence intervals resampled at the complete-movie level.

S7 Prevalence

S8 PPI correction

Prediction-powered corrections preserve the collection’s broad trends while widening confidence sets; comparisons show similar temporal dynamics across subcollections and report how women’s attention varies by shot-distance category.

S8 PPI correction

S9 Comparison by subcollection

S10 Gender differentials by category

S11 Animated gender

The supplementary analysis compares gendered shot-distance prevalence for animated characters and live-action characters. Its tables report frame shares and movie-resampled confidence intervals, and a figure places animated results alongside live-action.

Animated characters

  • Tables S3 and S4 report shot-distance frame shares for men and women among animated human and non-human characters, with movie-resampled 95% bootstrap confidence intervals.The samples are year-matched to animated/live-action comparisons.Source: Table S3 · p. 31, Table S4 · p. 31

Live-action comparison samples

  • Table S5 reports gendered shot-distance frame shares for live-action characters in the year-matched sample, while Table S6 covers popular films across all years.Both tables report movie-resampled 95% bootstrap confidence intervals.Source: Table S5 · p. 32, Table S6 · p. 32

Animated/live-action comparison

  • Figure S20 places the paired animated-character results alongside live-action results for comparison.The paired test distinguishes animated human and non-human characters from live-action characters.Source: S11 Animated gender · p. 33, Figure S20 · p. 33
Loading 2609.31734v1…