Features, guides, and ideas from the Labelme team.
A shape label drawn on the canvas now carries its group ID, the same way the Shape List writes it. Two shapes in group 1 both read "truck (1)"; a shape with no group ID still reads "truck".
Read articleThe Brightness/Contrast dialog labels each slider as a percentage. 100% is the image as it loaded, 150% is half again as bright, so you can read an adjustment and repeat it.
v7.3.0 renames three View actions, the two clipboard actions, and the Annotation List dock. Shortcuts and behavior are unchanged, so this is the old-to-new map.
Rectangle, oriented rectangle, polygon, mask, line, or point. Three questions that settle the choice before you label a thousand images, and what each shape is worth once it reaches the exporter.
Adjust Polygon Detail while an AI polygon is still a live preview. Move left for a cleaner outline or right to follow finer edges.
Choose Automatic, Uniform, or By Label from Settings. The canvas, Annotation List, and Label List update immediately.
Ctrl/Cmd-wheel zoom now keeps the image point beneath your cursor fixed, so you can inspect a detail without losing it as the canvas changes size.
Undo now activates after the first shape on a raw image. It removes a bad start from the canvas and Annotation List; auto-save updates the annotation file immediately.
Typing in File List search no longer rescans your folder or interrupts unsaved work. Filter the list while your current image and annotation stay on the canvas.
When the next image or annotation cannot load, your current image, shapes, canvas, and File List selection stay active instead of disappearing.
Six toggles used to switch themselves off every time you reopened Labelme. They now stay where you put them, and the Settings dialog covers 19 preferences instead of 8.
An AI prediction that overlapped a Shape you already had used to disappear. Predictions now become new Shapes, an optional setting highlights the matched Shape instead, and point prompts steer which candidate you get.
If Labelme cannot finish saving, your last successful annotation stays intact so you can fix the problem and try again.
Saving to a second Windows drive no longer closes Labelme, and the File List stops listing an image twice or leaving a dialog-opened image unchecked.
One review loop to learn first, two Labelme keyboard shortcuts worth understanding, and a printable reference for every default binding.
A recent active-learning study spent 20% of a fixed annotation budget on checking existing labels. Its result makes a strong case for planning review before a project starts.
Labelme v7.0.4 fixes an AI Assist / AI Box crash that hit on older Pillow versions when a new detection overlapped an existing polygon or oriented-rectangle shape.
Labelme v7.0.3 fixes a crash where deleting an annotation from the list's right-click menu brought the app down on your next shape. It also stops AI polygon output from choking on very thin detections.
Labelme v7 adds a color theme setting with System, Light, and Dark. Dark mode is new, System tracks your OS appearance, and the switch applies live without a restart.
A BETA preview in Labelme v7. Place polygon and shape points beyond the image boundary, so objects that run off the frame keep their true geometry instead of being clamped to the edge.
A BETA preview in Labelme v7. Draw each shape's label text directly on the canvas so you can read what's what without cross-referencing the polygon list.
Labelme v7 replaces the open-the-YAML-in-an-editor preference flow with a Settings dialog that applies changes live, preserves your comments, and lets you pick the language in-app.
The toolkit now converts oriented bounding boxes both ways. Export a Labelme dataset to YOLO-OBB for YOLOv8-OBB training, and import a YOLO-OBB dataset back into editable Labelme shapes.
A three-click sequence used to crash Labelme whenever you changed draw modes mid-shape. v6.3.1 fixes it, and stops empty shapes (zero-area rectangles, zero-length lines, single-point polygons) from getting saved into your annotation files.
Two suppression passes in Labelme v6.3 stop AI-Box, AI-Points, and AI-Text from stacking redundant predictions on the same object, whether they came from nested SAM granularities or from regions you've already labeled.
Pick the output shape for AI-Points and AI-Box: polygon, mask, rectangle, oriented rectangle, or circle. Fitted to the SAM mask, no post-export conversion.
Select many annotations at once and apply the same edit to all of them. Ctrl/Cmd+A grabs every shape on the image, range-select works in the label list, and hide/show propagates across the whole selection as one undoable step.
A new shape type in Labelme v6.2. Three clicks to draw, corner drag to resize, edge midpoint to rotate, with an arrow on the canvas so the heading is never ambiguous.
A podcast with Robin Cole on annotating satellite imagery in Labelme, plus an end-to-end NAIP to QGIS demo.
The first AI click of a fresh install used to freeze the app for minutes while the model downloaded in silence. v6.1 shows a progress bar with bytes, ETA, and a cancel button.
Labelme v6.1 collapses four near-identical AI buttons into AI-Points and AI-Box, each with a polygon-or-mask output toggle. A separator between manual and AI tools makes the right button easier to find when you haven't opened the app in a week.
Drag a single bounding box around a cluster and SAM3 returns one shape per instance inside it. Built for parking lots, shelves, microscopy, and anything else where every visible object is the same kind of thing.
Labelme v6 stops base64-embedding the image inside every annotation JSON by default. Files get an order of magnitude smaller, diffs become readable, and version control stops choking on the binary blob.
Labelme v6 saves annotations as you work. The most common way to lose hours of labeling — forgetting to Ctrl+S before closing — is gone for new installs.
Labelme v6 opens multi-gigapixel images and float32 GeoTIFFs that v5 refused to load. Pan, zoom, and annotate a 26,000 x 22,000 px aerial tile without the app locking up.
Annotate a custom dataset with Labelme, export to YOLO format with labelmetk, and train with Ultralytics. All offline.
AI tools can propose annotations automatically. Here's how to review them fast with Labelme: fix what's wrong, skip what's right.
Cloud annotation tools upload your images to their servers. For medical AI teams, that's a HIPAA problem. Labelme downloads its AI models once, then runs entirely offline, so your images never leave your machine.
Cloud annotation tools send your training data to third-party servers. Here's why that's a growing problem, and how offline-first tools like Labelme solve it.
