GeoReason-Bench #135
Description
📌 Overview
Total Samples: 4,000 high-fidelity reasoning trajectories
Dataset Size: 90.3KB (only json, no images)
Source (): DOTA, DIOR
📌 Data Pipeline
The original annotations were derived from geometric primitives (e.g., scale, orientation) extracted from the DOTA and DIOR datasets. These primitives were aggregated into morphological patterns (e.g., spatial density, object spacing, clustering configurations) and combined with global scene descriptions generated by advanced Vision-Language Models (VLMs). By integrating structural metadata with semantic context, GPT-4o was utilized to synthesize the reasoning trajectories and final answers.