Skip to content

BMVC 2026

PERSISTPersistent-State Discrimination
for Shot Boundary Detection

Tingyu Lin1, Christian Stippel1, Armin Dadras1,2, Jakob Zenzmaier3, Florian Kleber1, Wolfgang Aigner2, Robert Sablatnig1

1 Computer Vision Lab, TU Wien

2 Institute of Creative\Media/Technologies, St. Pölten University of Applied Sciences

3 Ludwig Boltzmann Institute for Digital History

The difference is in what follows

See what persists.

The original figure compares four real clips and their recorded boundary scores.

Real clips with recorded frame scores. The cue-only baseline uses the same backbone and training setup. Original figure ↗

How it works

Learning a persistence criterion.

Local change provides evidence. A continuous latent state helps distinguish lasting transitions from changes that soon return.

Full architecture ↗

During training, PERSIST learns whether a visual change lasts or the video returns to its earlier appearance. This helps it distinguish new shots from brief disturbances.

Evaluation

What changes in the prediction?

The model learns to separate transient disturbances from real cuts, while remaining competitive across four video domains.

Across four benchmarks

Published F1 at a shared threshold of 0.5.

Shot boundary detection F1 at threshold 0.5 across four datasets. Best result in each column is bold.
MethodClipShotsBBCAutoShotHIST-GST12
Pixel difference0.03510.00740.06830.0000
HSV histogram0.57370.77860.67920.1544
Edge change ratio0.35100.55520.35050.6019
DeepSBD · AlexNet0.68430.82020.75390.8535
DeepSBD · ResNet180.75550.86220.73030.8611
TransNetV20.77930.96110.79930.9204
PERSIST0.80230.97000.79530.9264
Best per column in bold. Full protocols and oracle results ↗

Similar evidence. Different scores.

Real cuts remain; pseudo-events are suppressed.

Hard cutsHigh-evidence pseudo-events
Hard cuts and pseudo-events have similar mean local evidence, 0.865 and 0.841, but different final scores, 0.725 and 0.269.
Mean scores on ClipShots. Analysis in the paper ↗

A closer look

From false alarms to frame-level decisions.

Compare the detections, then follow the evidence through two video examples.

Equal sensitivity

Fewer false alarms at equal recall

Keeping sensitivity to real transitions exposes the difference in false detections.

False-positive counts for PERSIST and TransNetV2 on rendered diagnostic clips and natural ClipShots footage at matched recall.
False positives at matched true-transition recall, evaluated separately on each split.

Across failure modes

Five kinds of disturbance

The reduction spans flashing light, rapid camera motion, overlays and damaged film.

False-positive detections for flash, fast pan, text overlay, archival degradation and scratches, comparing PERSIST with TransNetV2.
False-positive detections at matched true-transition recall across five rendered disturbance types.

Frame by frame

A cut and a passing event

Read the footage alongside the model’s local evidence and final score.

Published true-cut and pseudo-event examples with corresponding frame strips, local evidence and final-score curves.
Two published examples: the final score follows a real cut and stays low during a passing disturbance.

A geometric view

Reading suppression through time

Track the gap between local evidence and the final score through both examples.

The published space-time cube, with filmstrip frames and trajectories showing the gap between local evidence and final score.
Trajectory height encodes local evidence minus final score; the filmstrips show the corresponding video frames.

Further analysis in the paper and supplementary material ↗

Citation

Cite PERSIST

@inproceedings{lin2026persist,
  author    = {Lin, Tingyu and Stippel, Christian and Dadras, Armin and
               Zenzmaier, Jakob and Kleber, Florian and Aigner, Wolfgang and
               Sablatnig, Robert},
  title     = {{PERSIST}: Persistent-State Discrimination for Shot Boundary Detection},
  booktitle = {Proceedings of the British Machine Vision Conference (BMVC)},
  year      = {2026}
}

Questions about the work? Contact Tingyu Lin.