Embed the world: Multimodal AI for searchable aerial imagery at scale
TL;DR
AWS and Vexcel describe a pipeline that makes aerial imagery searchable with natural language: multimodal embeddings and optional captions via Amazon Bedrock, with search handled by Amazon OpenSearch Serverless. The evaluation used Grant Park in Chicago, OpenStreetMap as ground truth, and about 100 configurations across two query types: swimming pools as discrete objects and roads as distributed infrastructure.
Nauti's Take
The post is obviously AWS-heavy, but the useful takeaways are specific. Captions are not a decorative add-on here; they materially improve retrieval when paired with image embeddings.
The experiments also show why generic AI search gets oversold: pools, roads, building height, and complex location queries need different strategies. The real lesson is simple: build the benchmark first, then swap models.
Briefingshow
The interesting part is not the AWS product wrapper, but the evaluation design. Aerial imagery is not a normal photo archive: one place has multiple views, and one correct tile can mean something different from finding every real object. Teams building this kind of system need measurement before interface polish.
Sources
