Home / Current Issue / Paper 1722842
Discontinuity Is Not a Cut: Automatic Shot Boundary Detection and The Measurement of Film Editing
Subject area: Arts, Social Sciences and Humanities · Area of research: Film and Television
Abstract
Film studies increasingly computes average shot length from automatically detected shot boundaries. Detectors do not identify shots. They identify discontinuities in the visual signal, which coincide with shots only approximately. Because that divergence is structured by the footage rather than random, it biases shot-length statistics rather than adding noise. Using the ClipShots ground truth (5,138 shots across 92 videos) with detector performance figures published for that corpus, we simulate how detection error propagates into the statistics the field reports. In these simulations average shot length is systematically deflated, by 15.4% and 10.6% for the two stronger detector profiles, because false boundaries outnumber missed ones. Video is made to appear as though it cuts faster than it does. The bias does not track detector quality, and it is confounded with the object of study. Since gradual transitions are missed far more often than cuts, a film's bias varies almost perfectly with the proportion of its transitions that are gradual, itself a stylistic property. Films that cut hard and films that dissolve are distorted in opposite directions by the feature distinguishing them. Differences below about 5% prove unrecoverable under every profile tested. Researchers comparing moving images on shot length cannot assume the size of their measurement error, nor remove it by choosing a detector with a better F-score.
Keywords
average shot length, cinemetrics, film editing, measurement error, shot boundary detection
How to cite this paper
@article{1722842,
author = {Md Riaz Uddin, Sameer Ahmed},
title = {Discontinuity Is Not a Cut: Automatic Shot Boundary Detection and The Measurement of Film Editing},
journal = {Iconic Research And Engineering Journals},
year = {2026},
volume = {9},
number = {12},
pages = {3972-3986},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1722842.pdf},
abstract = {Film studies increasingly computes average shot length from automatically detected shot boundaries. Detectors do not identify shots. They identify discontinuities in the visual signal, which coincide with shots only approximately. Because that divergence is structured by the footage rather than random, it biases shot-length statistics rather than adding noise. Using the ClipShots ground truth (5,138 shots across 92 videos) with detector performance figures published for that corpus, we simulate how detection error propagates into the statistics the field reports. In these simulations average shot length is systematically deflated, by 15.4% and 10.6% for the two stronger detector profiles, because false boundaries outnumber missed ones. Video is made to appear as though it cuts faster than it does. The bias does not track detector quality, and it is confounded with the object of study. Since gradual transitions are missed far more often than cuts, a film's bias varies almost perfectly with the proportion of its transitions that are gradual, itself a stylistic property. Films that cut hard and films that dissolve are distorted in opposite directions by the feature distinguishing them. Differences below about 5% prove unrecoverable under every profile tested. Researchers comparing moving images on shot length cannot assume the size of their measurement error, nor remove it by choosing a detector with a better F-score.},
keywords = {average shot length, cinemetrics, film editing, measurement error, shot boundary detection},
month = {June},
}