Benchmarking Composed Image Retrieval for Applied Earth Observation
Researchers have developed a new benchmark for Composed Image Retrieval (CIR) specifically tailored for Earth Observation (EO) data. This benchmark evaluates existing CIR methods on satellite imagery, revealing that training-free composition techniques offer robust baselines. The study also introduces a new dataset, xView2-CIR, focused on disaster monitoring, highlighting unique challenges in change-centric retrieval that differ from attribute-based methods. AI
IMPACT Establishes a standardized benchmark for satellite image retrieval, potentially improving disaster monitoring and archive exploration tools.