GeoHeight-Bench - Instance Extraction: leaderboard
Metric: Text accuracy (%) on the Instance Extraction task (the model gives an instance of a named category and its height), on GeoHeight-Bench (object height above ground from nDSM products; GeoNRW, FLAIR and RSMSS aerial tiles); zero-shot on RGB imagery only; Qwen2.5-1.5B extracts template answers, which count when continuous values fall within 20 percent relative error, and Qwen2.5-7B judges open-ended answers for semantic and factual consistency; higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 29.82 | #237 |
| 2 | InternVL2-8B | 13.51 | #826 |
| 3 | GPT-4o | 12.2 | #333 |
| 4 | Pixtral-12B | 9.71 | #795 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=geoheight-bench-instance-extraction · How It Works · Data refreshed daily, snapshot 2026-10-11.