ThermEval - Human Counting (FLIR): leaderboard
Metric: Mean absolute error (people, 0 or more) on Task 3, the number of people in 10,000 grayscale thermal FLIR-ADAS images; zero-shot with fixed one-word or one-number prompts, a single forward pass; Gemini 2.5 models parse free-form answers into the label or number only when the output deviates from the format; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.2 11B Instruct | 2.84 | #1112 |
| 2 | InternVL3-38B | 2.93 | #395 |
| 3 | InternVL3-14B | 2.99 | #494 |
| 4 | InternVL3-8B | 3.02 | #606 |
| 5 | Qwen 2.5 VL 7B Instruct | 3.55 | #643 |
| 6 | Qwen 2.5 VL 32B Instruct | 3.59 | #443 |
| 7 | Qwen 2 VL 7B Instruct | 3.66 | #816 |
| 8 | MiniCPM-V-2.6 | 3.88 | #825 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=thermeval-human-counting-flir · How It Works · Data refreshed daily, snapshot 2026-10-11.