What it does RectLabel is an offline image annotation tool for object detection and segmentation, distributed as a paid Mac app with a free tier. This repository is its public support channel—issues, feature requests, and documentation—rather than open-source code. The README functions as a feature list and marketing page for the App Store downloads.

The interesting bit The app integrates a surprising breadth of models locally: Segment Anything 2 and 3 for polygon and pixel labeling, Cellpose for biological segmentation, Core ML models including RF-DETR and YOLO26 for auto-labeling, plus tracking and OCR. That’s a lot of neural network inference running offline on Apple hardware, not a web service with usage caps.

Key highlights

• Supports bounding boxes, polygons, pixels, oriented bounding boxes, cubic Bézier curves, keypoints with skeletons, and point annotations
• Auto-labeling via SAM 3 text/box prompts, SAM 2, Cellpose, Core ML detection models, and a tracking model
• Exports to YOLO, COCO, CreateML, and DOTA formats; also exports indexed color and grayscale masks
• Includes video frame extraction, image augmentation, gallery search across objects/attributes/filenames/memos, and configurable hotkeys
• Two App Store tiers: RectLabel and RectLabel Pro (pricing not listed in repo)

Caveats

• The app is macOS-only and distributed through the App Store, not open source; this repo contains no code to inspect or build from
• “Jupyter Notebook” as the listed repository language is misleading—there appear to be no notebooks in the README
• Several listed features (YOLO26, RF-DETR) are mentioned without version specificity or performance claims

Verdict Mac-owning ML engineers who need offline, format-flexible annotation and want SAM-powered auto-labeling without cloud dependencies should evaluate the free tier. Windows and Linux users, open-source purists, and anyone wanting to self-host or modify the tool are out of luck.