The 16K-star labeler that still runs on your laptop

What it does

Labelme is a desktop GUI for drawing polygons, rectangles, circles, lines, and points on images to build training datasets for computer-vision models. It exports to Pascal VOC and COCO formats, handles video frames, and supports whole-image flag annotations for classification tasks. Everything saves to local JSON files.

The interesting bit

The project started as a Python/Qt clone of MIT’s original LabelMe, but the maintainers resisted the obvious SaaS pivot. Instead they bolted on local AI assistance—SAM and EfficientSAM for point-to-mask generation, YOLO-World and SAM3 for text-prompted annotations—while keeping the tool runnable offline. There’s even a paid standalone executable for users who want to skip the Python/Qt dependency dance entirely.

Key highlights

• Native AI assistance via SAM/EfficientSAM point-to-polygon and YOLO-World/SAM3 text-to-annotation
• Exports Pascal VOC (semantic/instance) and COCO (instance) formats without conversion scripts
• 20-language UI including CJK, Arabic, and Persian
• Video annotation support with frame navigation
• Standalone app available via one-time payment at labelme.io

Caveats

• Building the standalone executable currently requires pinning numpy<2.0 due to unresolved build errors
• The AI model integration is relatively new; the README doesn’t specify hardware requirements or inference speeds

Verdict

Worth a look if you need a free, offline-first annotation tool with modern AI shortcuts and don’t mind Qt’s aesthetic. Skip it if you’re already embedded in a collaborative cloud labeling pipeline or need enterprise review workflows.