HorusEye: Language as Dynamic Attention for Emergency Visual Analysis
A new research paper introduces HorusEye, a framework designed for emergency visual analysis that treats language as dynamic attention. The study benchmarks various vision-language models (VLMs) like Gemini, Qwen2-VL, BLIP-2, LLaVA, and Kosmos-2 on a degraded dataset simulating conditions such as fog, smoke, and thermal imagery. Key findings indicate that language feedback significantly impacts model performance differently across VLMs, with Gemini showing substantial improvement in thermal conditions while Qwen2-VL degrades. The research also highlights a 'Thermal Paradox' where image cropping strategies effective for RGB fail in thermal imagery, and notes that BLIP-2 uniquely hallucinates more under degradation. AI
IMPACT Introduces a novel approach for emergency visual analysis, highlighting model-specific performance variations and challenges in degraded conditions.