ghost233lism/GeoAgent
GeoAgent is a 7 billion parameter vision-language model developed by Modi Jin et al. from VCIP, Nankai University, and Northwestern Polytechnical University, built upon Qwen2.5-VL-7B-Instruct with a 32768-token context length. It is specifically designed for image geolocation, reasoning with human-like precision to derive fine-grained address conclusions across various geographic grains. The model's primary differentiator is its use of Geo-similarity and Consistency rewards, trained on the novel GeoSeek dataset, enabling interpretable chain-of-thought reasoning for accurate global geolocation.
Loading preview...
GeoAgent: Vision-Language Model for Image Geolocation
GeoAgent is a 7 billion parameter vision-language model (VLM) developed by researchers from Nankai University and Northwestern Polytechnical University, specifically engineered for image geolocation. Built upon the Qwen2.5-VL-7B-Instruct architecture, GeoAgent excels at identifying geographic locations from images with human-like reasoning, providing fine-grained address conclusions across city, region, country, and continent levels.
Key Capabilities
- Precise Image Geolocation: Accurately determines geographic locations from visual input.
- Interpretable Chain-of-Thought (CoT) Reasoning: Generates clear, step-by-step reasoning processes for its geolocation predictions.
- Enhanced Reward Mechanisms: Incorporates a novel Geo-similarity reward (combining spatial and semantic similarity) and a Consistency reward (ensuring reasoning integrity) to improve accuracy and robustness.
- Trained on GeoSeek Dataset: Utilizes a specialized dataset, GeoSeek, which includes human-annotated CoT data and bias-reduced sampling for comprehensive training.
Good for
- Geographic Information Systems (GIS): Applications requiring automated, precise image-based location identification.
- Environmental Monitoring: Analyzing images to pinpoint specific environmental features or changes globally.
- Autonomous Navigation: Enhancing systems that require visual context for location awareness.
- Research in Vision-Language Understanding: Exploring advanced reasoning and multi-modal integration for complex tasks like geolocation.