AeroGround - Region Counting: leaderboard
AeroGround (2026) tests vision-language models on paired UAV and ground-level views of simulated open environments. Exact-match accuracy in percent on closed-form questions; higher is better. The model counts the instances of a category inside a region marked in the aerial view.
Metric: Accuracy (%; Spatial Understanding: count category instances inside a marked aerial region; aerial-ground collaborative questions from paired simulated UAV and ground views, exact-match scoring of the parsed option or mapping, invalid output scored wrong; one deterministic run). Source: arxiv.org. Saturation forecast: Not forecast. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B (Thinking) | 45.3 | #557 |
| 2 | GPT-5.4 | 39.1 | #121 |
| 3 | Qwen 3 VL 8B (Thinking) | 39.1 | #887 |
| 4 | Kimi K2.5 | 37.5 | #252 |
| 5 | Qwen 3 VL 8B Instruct | 34.4 | #982 |
| 6 | Qwen 3 VL 32B Instruct | 34.4 | #582 |
| 7 | Gemini 3.1 Pro (Preview) | 23.4 | #91 |
| 8 | Seed 2.0 Pro | 20.3 | #209 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aeroground-region-counting · How It Works · Data refreshed daily, snapshot 2026-11-12.