Reviewing AI-generated annotations with LabelMe
SAM2, YOLO-World, and other generative pipelines can propose bounding boxes and masks automatically. They're not always correct. The point is to review them fast and fix what's wrong.
Accepting a correct annotation takes a keypress. Drawing one from scratch takes 30–60 seconds. You want to build the workflow around that gap.
Review loop
Install LabelMe and open your image directory. LabelMe's AI annotation generates proposals locally (SAM2 for polygons, YOLO-World for bounding boxes from text prompts). No internet needed after the initial model download.
Once proposals exist, the loop is: scan the annotations, delete the bad ones, adjust near-misses, press D, next image. For correct annotations you do nothing — just move on. See the starter guide for keyboard shortcuts.
On clean synthetic images with well-defined objects, this takes about 5–15 seconds per image.
Exporting
LabelMe saves annotations as JSON. Convert to training formats with the toolkit:
export-to-yolo— YOLO.txtformat (Ultralytics compatible)export-to-voc— Pascal VOC XML
Polygons are converted to bounding boxes automatically when exporting to rectangle-based formats.
Offline
Everything runs on your machine. Images, AI proposals, reviewed annotations — none of it leaves your disk. More on this in Why offline-first annotation matters.
In practice
With mostly correct AI proposals, 200–400 images reviewed per hour is realistic. Compare that to manual polygon annotation: 200–400 images per day.
The bottleneck is proposal quality. At 95% correct you're mostly pressing D. At 60% you're spending more time fixing than reviewing. Run a pilot batch of 50 images first to see where you land before committing to a full dataset.
LabelMe Pro includes SAM3 text-prompt annotation and the export toolkit, on top of SAM2. One-time purchase, $79.
LabelMe is an offline-first annotation tool with built-in AI.