EgoPolice: leaderboard
EgoPolice (Princeton University, 2026) tests zero-shot recognition of police-civilian actions in real police body-worn camera footage: five-option multiple choice over 12,000 questions on 1-second, 10-second and 1-minute clips. Accuracy in percent (chance 20%), higher is better.
Metric: Accuracy (%; overall over all 12,000 questions: 500 per answer class on 1-second and on 10-second clips and 200 per class on 1-minute clips; zero-shot five-option multiple choice (one visible police-civilian action, three actions absent from the clip and None of the above; chance 20%) on minimally edited COPA and Pasadena police body-worn camera clips; images at 1 fps (at least 8 frames) for image VLMs, author-recommended sampling for video VLMs, one shared prompt). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 77.2 | #266 |
| 2 | GPT-4.1 | 76.5 | #495 |
| 3 | Gemini 2.5 Flash | 67.5 | #453 |
| 4 | InternVL3-8B | 63.9 | #1688 |
| 5 | Qwen 2.5 VL 7B Instruct | 54.4 | #1833 |
| 6 | GPT-4.1 Nano | 17.3 | #1977 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=egopolice · How It Works · Data refreshed daily, snapshot 2026-11-13.