Coordinate Based Region Pointing And Gaze Detection

via “coordinate-based region pointing and gaze detection”

Tiny vision-language model for edge devices.

Unique: Region encoder subsystem directly outputs coordinate embeddings that map to pixel space, enabling end-to-end coordinate prediction without separate regression heads; coordinate transformations handle conversion between normalized and absolute coordinates, enabling flexible output formats.

vs others: Integrated into single model without separate pointing or gaze detection modules; enables spatial reasoning without training custom coordinate regression networks.