Import & export

Export formats

Format	Contains	Use it for
Labelme JSON	One JSON per image: boxes as rectangle, polygons, classes as flags	LabelMe and tools that read its format
YOLO boxes	Ready-to-train Ultralytics dataset: images/, labels/, data.yaml	YOLO detection training
COCO JSON	One coco.json with boxes, polygon segmentations, and areas	Detection and instance segmentation
Classification CSV	One row per image, one 0/1 column per class	Image classification
Native JSONL	Every annotation with its ID and status	Lossless backup and round-trips

Shapes that a format can't represent are counted and reported, not silently dropped: YOLO boxes skips polygons and classes, COCO skips classes, and the CSV skips shapes.

Output layout

Each export creates a new folder, so earlier exports are never overwritten:

Lannotex-export-1a2b3c4d/
├── manifest.json          # format, counts, skipped images and reasons
├── classes.json           # label list; YOLO class IDs follow this order
├── classification.csv
├── images.jsonl
├── annotations.jsonl
└── 00000001_street.json   # per-image files (Labelme JSON)

Images are referenced, not copied: exported records point to the original files. YOLO is the exception, described below.

What gets skipped

An image is left out of the export, and listed in manifest.json, if:

• its source file is missing,
• an annotation fails validation, or
• it has an unsaved recovery draft. Save it first.

Run View → Analysis or lannotex analyze beforehand to find these images.

YOLO: ready to train

The YOLO export is a complete Ultralytics dataset:

Lannotex-export-1a2b3c4d/
├── data.yaml              # path, train/val, class names
├── images/00000001_street.jpg
└── labels/00000001_street.txt   # class cx cy w h, normalized

yolo detect train data=Lannotex-export-1a2b3c4d/data.yaml model=yolo11n.pt

• Images are hard-linked into images/ when the export is on the same drive, so they take no extra space. Otherwise they're copied. Your originals are never modified.
• Images marked No target objects get an empty label file, so YOLO learns them as background.
• Images you haven't annotated yet are left out. If they were included, YOLO would treat them as "nothing here".
• train and val point at the same images. Split them yourself for a real evaluation.
• If you move the export folder, update path: in data.yaml.

Importing annotations

Project → Import annotations… accepts:

• a Labelme JSON file, or a folder of them (top level only), and
• a native JSONL export.

Each referenced image is checked: it must exist, and its dimensions must match the annotation file. Labels you don't have yet are created. An image that already has annotations or a draft in the project is skipped, never overwritten. Errors are listed per file.

See Coming from LabelMe for which LabelMe shape types are supported.

Headless export

lannotex export ./my-dataset ./exports --format "COCO JSON"

See Command line.