promptable segmentation

**Promptable Segmentation** is a **paradigm where segmentation masks are generated based on user inputs** — allowing users to interactively define what to cut out using points, bounding boxes, scribbles, or natural language, rather than relying on predefined fixed categories. **What Is Promptable Segmentation?** - **Definition**: Segmentation conditioned on external guidance (prompts). - **Shift**: Moves from "class-based" (segment all cars) to "instance-based" (segment *this* car). - **Interaction**: Often iterative; user clicks, model predicts, user corrects with more clicks. - **Flexibility**: Handles objects the model has never seen before (zero-shot). **Key Prompt Types** - **Spatial Prompts**: - **Points**: Foreground/background clicks. - **Boxes**: Bounding box around the object. - **Scribbles**: Rough lines drawn over the object. - **Semantic Prompts**: - **Text**: "Segment the red chair next to the window." - **Reference Image**: "Segment objects that look like this image." **Why It Matters** - **Annotation Speed**: Accelerates data labeling by 10-100x. - **Usability**: Makes powerful CV tools accessible to non-experts. - **Generalization**: Decouples "what" to segment from "how" to segment. **Promptable Segmentation** is **the interface for modern computer vision** — enabling dynamic human-AI collaboration for image editing, analysis, and content creation.