What problem does it solve?
Manually drawing object masks for video annotation is slow and labor-intensive. This Skill lets you click on any object in a video frame and instantly receive a precise segmentation mask, dramatically accelerating dataset labeling and annotation workflows.
Core Features & Use Cases
- Click-to-Segment: Send point or bounding-box prompts over a JSON-lines stdin/stdout protocol and receive masks with confidence scores and bounding boxes.
- Video Tracking: Segment an object in one frame and propagate the mask across subsequent frames in a clip.
- Hardware-Aware Setup: The deploy script auto-detects CUDA, Apple Silicon MPS, or CPU and installs the matching PyTorch build and SAM2 checkpoint.
- Use Case: In an annotation studio, a user clicks a person in a surveillance frame; the Skill returns a pixel mask and bounding box that can be saved as a labeled training sample for a detection model.
Quick Start
Run the deploy script to install dependencies and the model, then ask the assistant to segment the person in the current video frame using a click point.
