pith. sign in

arxiv: 2605.17095 · v1 · pith:WSSM4NDQnew · submitted 2026-05-16 · 💻 cs.CV · cs.AI· cs.LG

Visual Timelines of Police Encounters in Body-Worn Camera Footage: Operational Context and Activity Cataloging for Training and Analysis in OpenBWC

Angela Srbinovska , Christopher Homan , Adrian Martin , Ernest Fokou\'e This is my paper

Pith reviewed 2026-05-20 15:27 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords body-worn cameravideo window classificationoperational contextmotion intensityCLIP embeddingsoptical flowpolice encounter analysisvisual timelines
0
0 comments X

The pith

Body-worn camera footage can be turned into labeled 10-second windows that mark operational context and motion intensity shifts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a method to convert long body-worn camera videos into fixed 10-second windows that receive two labels: one for the operational context of the encounter and one for the level of physical motion intensity. Frames from each window are encoded with a CLIP model and combined with dense optical flow statistics, then fed to classifiers that reach 78.75 percent accuracy on context and 88.33 percent accuracy on activity in held-out test windows. The resulting time-aligned labels produce visual timelines that let analysts and trainers locate key moments without watching entire recordings. The approach incorporates a privacy-conscious labeling protocol and low-evidence flags for windows obscured by darkness, blur, or occlusion, along with integrity audits that demonstrate practical gains for incident review and officer training.

Core claim

The paper claims that BWC footage can be segmented into time-aligned 10-second windows, each labeled for operational context and motion intensity using a privacy-conscious protocol. Window representations formed by aggregating CLIP frame embeddings together with dense optical flow statistics allow classifiers to label context at 78.75 percent accuracy and activity at 88.33 percent accuracy on test data. These labeled sequences yield visual timelines that reduce the time required for full-video review and make training workflows more practical.

What carries the argument

Window-level representation formed by CLIP-encoded frames aggregated across sampled frames plus dense optical flow statistics, used to classify operational context and motion intensity.

If this is right

  • Analysts locate key encounter moments by scanning the timeline rather than watching complete videos.
  • Training sessions can focus directly on cataloged shifts in motion intensity and context.
  • Low-evidence flags automatically mark windows where visual content is unusable due to darkness, blur, or occlusion.
  • Integrity audits confirm that the labeled timelines support faster incident review and training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same windowing and dual-label approach could be applied to other long-form surveillance video domains where context and intensity matter.
  • Adding temporal modeling across adjacent windows might further improve detection of activity transitions.

Load-bearing premise

Human-provided labels for operational context and motion intensity are sufficiently consistent and CLIP embeddings plus optical-flow statistics capture the necessary visual cues even when footage contains darkness, blur, or occlusion.

What would settle it

Running the trained classifiers on a fresh collection of body-worn camera videos that carry independently generated labels and measuring whether context accuracy falls below 70 percent or activity accuracy falls below 80 percent would test the central performance claim.

Figures

Figures reproduced from arXiv: 2605.17095 by Adrian Martin, Angela Srbinovska, Christopher Homan, Ernest Fokou\'e.

Figure 1
Figure 1. Figure 1: Most frequent transitions between adjacent labeled windows in the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Row-normalized agreement matrices comparing human annotator [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: GroundTruth window-level annotation interface (video preview is intentionally blurred in this paper for privacy reasons). For every 10-second window, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: End-to-End Visual Analysis Pipeline. and apply ℓ2 normalization z˜i,k = zi,k ∥zi,k∥2 . (2) This normalization has the effect of removing dependence on the magnitude of the embeddings and greatly increasing the stability of the pooled descriptor, Π ∈ {MEAN, MAX}, across sampled frames. To get a single descriptor per window, the normalized keyframe embeddings are then aggregated in a deterministic manner zi … view at source ↗
Figure 5
Figure 5. Figure 5: Operational context confusion matrix on the held-out clean test split [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Operational context confusion matrix on the held-out clean test split [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Activity confusion matrices for the strongest non-fused run (A6) run. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Most frequent recurring context confusions across experiments. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Most frequent recurring activity confusions across [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Recurring activity confusions across fused runs, shown as counts [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
read the original abstract

Law enforcement agencies are accumulating vast amounts of body-worn camera (BWC) footage. However, this remains operationally opaque. That is, analysts and trainers still have to invest considerable time watching full-length videos to pinpoint the start of key encounters and identify the points where activity shifts to something more physically intense. We present an approach to process BWC video into a time-aligned sequence of fixed-length 10-second windows, processed and labeled using a privacy-conscious protocol. Each window is labeled with two dimensions of information: (i) the operational context of the window and (ii) the level of motion intensity within the window, with low-evidence labels for windows for which insufficient evidence exists due to darkness, blur or occlusion. We train models to classify windows based on these two axes using frames sampled from each window encoded using CLIP model and aggregated into a window-level representation. We extract dense optical flow statistics for each window to capture motion intensity. On test windows the best context model achieves 78.75% accuracy, and the best-accuracy activity model achieves 88.33%. We also included integrity audits to show the results and how the visual timeline representations support faster incident review and make the officer training workflow more practical.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes processing body-worn camera (BWC) footage into fixed-length 10-second windows that are labeled along two axes—operational context and motion intensity (with a low-evidence category for darkness, blur, or occlusion)—using a privacy-conscious protocol. Window-level representations are formed from CLIP-encoded frames and dense optical-flow statistics; supervised classifiers are trained on these features. On held-out test windows the best context model reaches 78.75 % accuracy and the best activity model reaches 88.33 % accuracy. The resulting visual timelines are presented as aids for faster incident review, officer training, and integrity audits.

Significance. If label quality can be demonstrated, the pipeline offers a practical, low-compute method for turning opaque BWC archives into searchable timelines. The choice of standard CLIP embeddings plus simple optical-flow aggregates keeps the approach deployable on modest hardware, and the explicit low-evidence label is a sensible safeguard. These strengths would support the claimed utility for training workflows once the reliability of the human targets is quantified.

major comments (1)
  1. [Labeling protocol section] Labeling protocol section: the paper introduces a low-evidence category for windows with insufficient visual information yet reports no inter-annotator agreement statistics (Cohen’s kappa, Fleiss’ kappa, or raw disagreement rates) on any subset of the labeled windows. Because every reported accuracy (78.75 % context, 88.33 % activity) is measured against these human labels, the absence of agreement metrics leaves open whether the classifiers are capturing stable visual patterns or annotator-specific noise; this directly affects the load-bearing claim that the timelines support reliable training and audit use.
minor comments (2)
  1. [Abstract and results] Abstract and results: the total number of windows, the train/test split sizes, and any cross-validation or ablation details on the low-evidence category are not stated, making it impossible to judge whether the quoted accuracies are robust.
  2. [Methods] Methods: the precise aggregation method used to turn per-frame CLIP embeddings into a single window vector (mean, max, attention, etc.) is not described; adding this detail would improve reproducibility.

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for the constructive review and for emphasizing the importance of demonstrating label reliability, which directly supports the practical claims of the work. We respond to the single major comment below.

read point-by-point responses

  1. Referee: [Labeling protocol section] Labeling protocol section: the paper introduces a low-evidence category for windows with insufficient visual information yet reports no inter-annotator agreement statistics (Cohen’s kappa, Fleiss’ kappa, or raw disagreement rates) on any subset of the labeled windows. Because every reported accuracy (78.75 % context, 88.33 % activity) is measured against these human labels, the absence of agreement metrics leaves open whether the classifiers are capturing stable visual patterns or annotator-specific noise; this directly affects the load-bearing claim that the timelines support reliable training and audit use.

    Authors: We agree that the absence of inter-annotator agreement statistics is a limitation that should be addressed. Labeling was performed by a single trained annotator under a privacy-conscious protocol that restricted access to the sensitive BWC footage, so duplicate annotations were not collected and standard agreement metrics could not be computed. In the revised manuscript we will expand the Labeling Protocol section to describe the single-annotator process, explicitly state this as a limitation, and discuss its implications for interpreting the reported accuracies. We will also outline plans for future multi-annotator validation studies. These changes will clarify the nature of the human targets without overstating label stability. revision: yes

standing simulated objections not resolved
  • Quantitative inter-annotator agreement statistics (Cohen’s kappa, Fleiss’ kappa, or raw disagreement rates) cannot be provided because the labeling protocol used a single annotator and no duplicate labels were collected.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes a standard supervised classification pipeline: human annotators apply labels for operational context and motion intensity (including low-evidence categories) to fixed-length video windows, CLIP frame embeddings and optical-flow statistics are extracted as features, and models are trained then evaluated on held-out test windows to produce the reported accuracies. No equations, self-citations, or derivations are present that reduce any claimed result to a fitted parameter or input definition by construction; the outputs remain independent empirical measurements on external test data.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that CLIP embeddings transfer to operational-context classification in police footage and that a fixed 10-second window is an appropriate granularity; no new physical entities are postulated.

free parameters (1)
  • window_length = 10 seconds
    Fixed 10-second duration chosen for processing; directly affects label granularity and model input.
axioms (1)
  • domain assumption Pre-trained CLIP model yields embeddings informative for operational context in body-worn camera footage
    Invoked when frames are encoded with CLIP and aggregated for window-level classification.

pith-pipeline@v0.9.0 · 5774 in / 1376 out tokens · 66385 ms · 2026-05-20T15:27:36.911975+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Foundation/ArithmeticFromLogic.lean reality_from_one_distinction unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We train models to classify windows based on these two axes using frames sampled from each window encoded using CLIP model and aggregated into a window-level representation. We extract dense optical flow statistics for each window to capture motion intensity.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 2 internal anchors

  1. [1]

    Towards ai- driven policing: Interdisciplinary knowledge discovery from police body-worn camera footage,

    A. Srbinovska, A. Srbinovska, V . Senthil, A. Martin, J. McCluskey, J. Bateman, and E. Fokou ´e, “Towards ai- driven policing: Interdisciplinary knowledge discovery from police body-worn camera footage,” 2025, arXiv. [Online]. Available: https://arxiv.org/abs/2504.20007

    work page arXiv 2025

  2. [2]

    Police officer body-worn cameras: Assessing the evidence,

    M. D. White, “Police officer body-worn cameras: Assessing the evidence,” Office of Community Oriented Policing Services (COPS Office), U.S. Department of Justice, Washington, D.C., USA, Tech. Rep. NCJ 247941, 2014. [Online]. Available: https://www.ojp.gov/libr ary/publications/police-officer-body-worn-cameras-assessing-evidence

    work page 2014

  3. [3]

    The effect of police body-worn cameras on use of force and citizens’ complaints against the police: A randomized controlled trial,

    B. Ariel, W. A. Farrar, and A. Sutherland, “The effect of police body-worn cameras on use of force and citizens’ complaints against the police: A randomized controlled trial,”Journal of Quantitative Criminology, vol. 31, pp. 509–535, 2015. [Online]. Available: https://doi.org/10.1007/s10940-014-9236-3

    work page doi:10.1007/s10940-014-9236-3 2015

  4. [4]

    Video-based sso and body-camera data,

    J. McCluskey, C. D. Uchida, Y . Feys, and S. E. Solomon, “Video-based sso and body-camera data,” inSystematic Social Observation of the Police in the 21st Century. Cham: Springer, 2023, pp. 47–74. [Online]. Available: https://doi.org/10.1007/978-3-031-31482-7 4 10

    work page doi:10.1007/978-3-031-31482-7 2023

  5. [5]

    Terrill and L

    W. Terrill and L. Zimmerman, “Police use of force escalation and de-escalation: The use of systematic social observation with video footage police use of force escalation and de-escalation: The use of systematic social observation with video footage,”Police Quarterly, vol. 25, no. 2, pp. 155–177, 2022. [Online]. Available: https://nij.ojp.gov/library/publ...

    work page 2022

  6. [6]

    Systematic social observation of police-citizen encounters: Coding and measurement through body-worn cameras,

    R. E. Worden, B. P. Holladay, S. J. McLean, H. Cochran, and D. L. Reynolds, “Systematic social observation of police-citizen encounters: Coding and measurement through body-worn cameras,” Justice Quarterly, vol. 42, no. 7, pp. 1410–1443, 2025. [Online]. Available: https://doi.org/10.1080/07418825.2025.2463406

    work page doi:10.1080/07418825.2025.2463406 2025

  7. [7]

    Baselining incivility in one- on-one police encounters from bwc archival footage: Exploratory study of race, gender and contact type effects,

    B. P. Holladay and D. A. Makin, “Baselining incivility in one- on-one police encounters from bwc archival footage: Exploratory study of race, gender and contact type effects,”Police Practice and Research, vol. 22, no. 6, pp. 1618–1636, 2021. [Online]. Available: https://doi.org/10.1080/15614263.2021.1914040

    work page doi:10.1080/15614263.2021.1914040 2021

  8. [8]

    Leveraging body-worn camera footage to assess the effects of training on officer communication during traffic stops,

    N. P. Camp, R. V oigt, M. G. Hamedani, D. Jurafsky, and J. L. Eberhardt, “Leveraging body-worn camera footage to assess the effects of training on officer communication during traffic stops,”PNAS Nexus, vol. 3, no. 9, 2024. [Online]. Available: https://doi.org/10.1093/pnasnexus/pgae359

    work page doi:10.1093/pnasnexus/pgae359 2024

  9. [9]

    Body camera footage as data: Using natural language processing to monitor policing at scale & in depth,

    N. P. Camp and R. V oigt, “Body camera footage as data: Using natural language processing to monitor policing at scale & in depth,” Behavioral Science & Policy, vol. 10, no. 2, pp. 16–25, 2024. [Online]. Available: https://doi.org/10.1177/23794607241308636

    work page doi:10.1177/23794607241308636 2024

  10. [10]

    Cvat-bwv: A web-based video annotation platform for police body-worn video,

    P. Hejabi, A. K. Padte, P. Golazizian, R. Hebbar, J. Trager, G. Chochlakis, A. Kommineni, E. Graeden, S. Narayanan, B. A. T. Graham, and M. Dehghani, “Cvat-bwv: A web-based video annotation platform for police body-worn video,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, 2024, pp. 8674–8678. [Onli...

    work page doi:10.24963/ijc 2024

  11. [11]

    Improving Calibration of Thermal Stereo Cameras Using Heated Calibration Board

    H. Sameer, J.-L. Dugelay, M. Rizal Mohd Isa, and M. Adib Khairuddin, “Moving forward with bwc: The faleb dataset for multimodal image analysis,” inIEEE International Conference on Image Processing, 2025, pp. 1552–1557. [Online]. Available: https://doi.org/10.1109/ICIP 55913.2025.11084403

    work page doi:10.1109/icip 2025

  12. [12]

    Video analysis for body-worn cameras in law enforcement,

    J. Corso, A. Alahi, K. Grauman, G. D. Hager, L.-P. Morency, H. Sawhney, and Y . Sheikh, “Video analysis for body-worn cameras in law enforcement,” Computing Community Consortium, Washington, D.C., USA, Tech. Rep. White Paper 11, 2015. [Online]. Available: https://cra.org/ccc/wp-content/uploads/sites/2/2015/01/CCCWhitepaper onBodyCamerasinLawEnforcement.pdf

    work page 2015

  13. [13]

    Semi-Supervised First-Person Activity Recognition in Body-Worn Video

    H. Chen, H. Li, A. Song, M. Haberland, O. Akar, A. Dhillon, T. Zhou, A. L. Bertozzi, and P. J. Brantingham, “Semi-supervised first-person activity recognition in body-worn video,” 2019, arXiv. [Online]. Available: https://arxiv.org/abs/1904.09062

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  14. [14]

    Ego-motion classification for body-worn videos,

    Z. Meng, J. S ´anchez, J.-M. Morel, A. L. Bertozzi, and J. P. Brantingham, “Ego-motion classification for body-worn videos,” inImaging, Vision and Learning Based on Optimization and PDEs. Cham: Springer, 2018, pp. 221–239. [Online]. Available: https://doi.org/10.1007/978-3-319-91274-5 10

    work page doi:10.1007/978-3-319-91274-5 2018

  15. [15]

    Temporal segmentation of egocentric videos,

    Y . Poleg, C. Arora, and S. Peleg, “Temporal segmentation of egocentric videos,” inIEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 2537–2544. [Online]. Available: https://doi.org/10.1109/CVPR.2014.325

    work page doi:10.1109/cvpr.2014.325 2014

  16. [16]

    Foundation models for video understanding: A survey,

    N. Madan, A. Møgelmose, R. Modi, Y . S. Rawat, and T. B. Moeslund, “Foundation models for video understanding: A survey,” 2024, arXiv. [Online]. Available: https://arxiv.org/abs/2405.03770

    work page arXiv 2024

  17. [17]

    Internvideo2: Scaling foundation models for multimodal video understanding,

    Y . Wang, K. Li, X. Li, J. Yu, Y . He, C. Wang, G. Chen, B. Pei, Z. Yan, R. Zhenget al., “Internvideo2: Scaling foundation models for multimodal video understanding,” inComputer Vision – ECCV 2024. Cham: Springer Nature Switzerland, 2024, pp. 396–416. [Online]. Available: https://link.springer.com/chapter/10.1007/978-3-031-73013 -9 23

    work page doi:10.1007/978-3-031-73013 2024

  18. [18]

    Zelda: Video analytics using vision-language models,

    F. Romero, C. Winston, J. Hauswald, M. Zaharia, and C. Kozyrakis, “Zelda: Video analytics using vision-language models,” 2023, arXiv. [Online]. Available: https://arxiv.org/abs/2305.03785

    work page arXiv 2023

  19. [19]

    Learning Transferable Visual Models From Natural Language Supervision

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” 2021, arXiv. [Online]. Available: https: //arxiv.org/abs/2103.00020

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  20. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    M. Cherti, R. Beaumont, R. Wightman, M. Wortsman, G. Ilharco, C. Gordon, C. Schuhmann, L. Schmidt, and J. Jitsev, “Reproducible scaling laws for contrastive language-image learning,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 2818–2829. [Online]. Available: https://doi.org/10.1 109/CVPR52729.2023.00276

    work page arXiv 2023

  21. [21]

    Research computing services,

    R. I. of Technology, “Research computing services,” 2026. [Online]. Available: https://www.rit.edu/researchcomputing/

    work page 2026

  22. [22]

    Two-frame motion estimation based on polynomial expansion,

    G. Farneb ¨ack, “Two-frame motion estimation based on polynomial expansion,” inImage Analysis, vol. 2749. Berlin, Heidelberg: Springer, 2003, pp. 363–370. [Online]. Available: https://doi.org/10.1007/3-540 -45103-X 50 11 APPENDIXA INFERENCE ANDVISUALTIMELINECONSTRUCTION Algorithm 1: Visual Timeline Generator Input :Incident videosV; window lengthL=10; fra...

    work page doi:10.1007/3-540 2003