Text search over viewer-captioned snapshots from two fixed underwater cameras at Utopia Village, Roatán. Each query is answered three ways, so you can see what a few thousand captions teach a frozen image-text model about one camera.
SigLIP2, frozen
The base model as released (base/16, 256 px). No camera-specific training.
Two linear maps
One linear map on the query embedding and one on the image embedding, fitted to this camera's captions.
Two maps over sets
The same two maps applied to every phrase of the query and every patch of the image; each phrase is matched to its best patch.