Segment Anything
Meta released a model that isolates any object in an image from a point or box prompt, together with a dataset of a billion masks.
Why it matters
The promptable foundation model idea crossed from language into vision, and segmentation stopped needing training for each task.
Before this, every application area, medical imaging, satellite data, industrial inspection, trained its own segmentation model on its own annotations. SAM works without fine-tuning on domains it has never seen. The SA-1B dataset was assembled semi-automatically: the model annotated, people corrected, the model retrained. The weights were released openly.