EgoPolice: A Benchmark for Egocentric Video Understanding in High-Stakes Police Body-Worn Camera Footage

Jul 16, 2026

Upload by Sage

EgoPolice introduces a carefully curated dataset of real, egocentric police–civilian interactions drawn from publicly available body-worn camera videos. Police-civilian action labels are selected for relevance to behavioral research and annotated with second-by-second granularity. The footage is challenging due to rapid and irregular camera motion, dense human contact, and rare high-stakes events. The benchmark offers classification and multiple-choice QA tasks, evaluates open- and closed-source models, and motivates scalable event detection for efficient downstream human review.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages36

Reading Time1 h 31 min

Introduction

Contributions

Related Works

arXiv :2607 .06468v 1 [ cs .CV] 7 Jul 2026 EgoPolice: A Benchmark for Egocentric Video Understanding in High-Stakes Police Body-Worn Camera Footage Max Gonzalez Saez-Diez⋆1 Jihoon Chung⋆1 Adam D. Wolsky 1 Gregory Lanzalotto2 Dean Knox2 Jonathan Mummolo 1 Brandon M. Stewart 1 Olga Russakovsky 1 1 Princeton University {saezdiez, jc5933, awolsky, jmummolo, bms4, [olgarus}@princeton.edu](olgarus}@princeton.edu) 2 University of Pennsylvania [glanza@wharton.upenn.edu](glanza@wharton.upenn.edu), [dcknox@upenn.edu](dcknox@upenn.edu) Abstract. We introduce EgoPolice, a carefully curated dataset of real, egocentric police–civilian interactions, sourced from publicly available body-worn camera videos. We select police-civilian action labels that are critical for police behavioral research and annotate them at a secondby-second granularity. The videos feature rapid and irregular camera motion, dense human interactions, and rare high-stakes events, making the dataset a challenging benchmark for motion-robust and contextaware egocentric perception. We provide two different tasks, classification and multiple-choice question-answering, and benchmark both opensource and closed-source models. We find that even the best video models like Gemini 2.5 Pro still struggle to accurately predict high-risk actions such as “Weapon Out”. Beyond serving as a benchmark, EgoPolice provides a foundation for developing models capable of identifying events of interest in large-scale body-worn camera video repositories, enabling more efficient downstream human review. Content Warning: This paper includes real police body-worn camera footage, including potentially distressing scenes. Keywords: Egocentric Dataset, Video Understanding, Police Body-Worn Camera Analysis 1 Introduction Recent advances in Vision-Language Models (VLMs) have generated significant interest in their application to a wide range of tasks and environments, including complex, high-stakes settings such as medical diagnosis, autonomous navigation, and disaster response [19, 50, 81] . Despite their potential to support real-world decision-making, failures of these models have exposed important ethical, legal, ⋆ Equal Contribution Project Website: [https://github.com/princetonvisualai/egopolice](https://github.com/princetonvisualai/egopolice) 2 Max Gonzalez Saez-Diez⋆ , Jihoon Chung⋆ et al. BWC Physical Interaction Any Officer Handcuffing BWC Weapon Out Civilian On Ground Hard Easy Fig. 1: Sample frames of EgoPolice. We show representative frames for Easy (Top, Cyan) versus Hard (Bottom, Red) samples. Easy samples are correctly classified by VideoMAE V2 and CLIP [62,78], while Hard samples are misclassified by both. Models perform well on clear viewpoints but struggle under low-light, occlusion, and distance. and safety concerns [6, 30, 55] . These are especially pronounced in the context of law enforcement, where errors can carry civil, legal, and safety consequences. The proliferation of Body-Worn Cameras (BWCs) has generated massive archives of footage, driving demand for automated analysis. However, the data captured by these cameras differs from standard training datasets as the videos are characterized by rapid camera motion, severe occlusions, and actions that are usually filtered out in academic datasets. While the failure to distinguish between an innocent gesture and a real threat can carry serious social consequences, commercial vendors [1, 7, 24, 60, 73] have entered the space with tools designed to interpret these records, even though the research community lacks the instruments to characterize their limitations and stress-test them against footage that is often difficult to parse even for a trained human observer (Figure 1) . To address these challenges, we introduce a carefully curated dataset of realworld, egocentric police–civilian interactions. Our work is motivated by the urgent need to benchmark and improve model reliability in first-pers

FAQs

What is EgoPolice and what data does it use?

EgoPolice is a curated dataset built from publicly available police body-worn camera videos showing real egocentric police–civilian interactions. It focuses on the dynamics and challenges of first-person footage in high-stakes contexts.

How are actions in EgoPolice labeled?

EgoPolice uses selected police-civilian action labels and annotates them at a second-by-second granularity. The labels target actions critical for police behavioral research.

What tasks and model benchmarking does EgoPolice support?

EgoPolice provides two tasks: classification and multiple-choice question-answering. It benchmarks both open-source and closed-source models and highlights difficulties predicting high-risk actions such as “Weapon Out” even for strong video models.

Jul 16, 2026

Upload by Sage