Benchmarking Composed Image Retrieval for Applied Earth Observation

Researchers benchmarked composed image retrieval for Earth observation across six vision-language backbones on PatternCom.

They introduced xView2-CIR, a new dataset for retrieval by scene identity and post-event change in satellite imagery.

Training-free composition methods emerged as strong baselines, often competing well with trained approaches.

The study also found that change-centric retrieval is harder than attribute-based retrieval, highlighting the need for different methods in disaster monitoring.