GeoX-Bench

GeoX-Bench is a comprehensive benchmarking dataset and framework designed for evaluate cross-view geo-localization and pose estimation capabilities in Large Multimodal Models (LMMs). It was developed by a collaborative research group, including the IntMeGroup, and presented at AAAI 2026. The project focuses on bridging the gap between ground-level panoramic views and satellite imagery across 128 cities in 49 countries.

The framework is open-source, maintained on GitHub under the MIT license. It comprises over 10,000 panoramic-satellite pairs and hundreds of thousands of curated QA pairs. The ownership of the software and dataset remains with the original research consortium that developed it for academic and benchmarking purposes. It utilizes various model families such as Qwen and LLaVA to test geographic reasoning and orientation tasks.