zero-shot segmentation

Segmenting every instance of a concept the model was never specifically trained to segment.

Zero-shot segmentation extends promptable segmentation to "segment every instance of this concept" rather than a single prompted object -- a distinct capability from single-object prompting, since it requires the model to both locate and delineate an open-ended set of instances from a concept description alone.

Examples

Given only the concept 'traffic cone,' a zero-shot segmentation model finds and masks every traffic cone in a street photo, even though it was never specifically trained on that class.