Annotate a single image

Unofficial preview. Docsbook assembled this page from the public wkentaro/labelme README.md and examples/ directory. It is not affiliated with, endorsed by, or maintained by the Labelme project. The official documentation lives at labelme.io/docs.

This is the single-image walkthrough from examples/tutorial. It takes one photo through annotation, visual inspection, and conversion into the files a training pipeline consumes. Run the commands from inside examples/tutorial in a clone of the repository.

Annotate the image#

labelme apc2016_obj3.jpg

Draw a shape around each object and give it a label. Labelme writes apc2016_obj3.json beside the image.

Labelme with several labelled polygons drawn over a shelf of objects

Check what you saved#

Reading raw JSON tells you little about whether the polygons landed where you meant. Draw them back over the image instead:

./draw_json.py apc2016_obj3.json

The saved annotation rendered back over the source image for visual checking

Convert to a dataset#

./export_json.py apc2016_obj3.json

This generates four standard files from the single JSON:

File	What it holds
`img.png`	The image itself
`label.png`	A uint8 label image, one class index per pixel
`label_viz.png`	A colourised visualisation of `label.png`
`label_names.txt`	The label name for each value in `label.png`

Inspect the label PNG#

label.png looks almost black when opened normally, because its pixel values are class indices such as 0, 1, 2, 3 rather than brightnesses. Render it with the helper script:

./draw_label_png.py apc2016_obj3/label.png

The label PNG rendered with a distinct colour per class index

Loading it in code has the same trap — use PIL.Image.open, described in reading annotations in Python.

Core concepts — what a shape, a label and a flag are
Bounding box detection — the same flow with rectangles
Annotation JSON reference — the file you just produced

Updated Aug 26, 2026