dataset-convert

name	dataset-convert
description	Convert datasets between COCO, LabelMe, and YOLO pose formats using Annolid-native converters.
metadata	{"annolid":{"requires":{"bins":[]}}}

Dataset Convert

Use this skill when the user asks to convert dataset formats.

Prefer the bundled converter script for deterministic runs:

python annolid/core/agent/skills/dataset-convert/scripts/convert_dataset.py ...

Supported conversions

1. labelme-to-coco: LabelMe directory -> COCO train/valid JSON.
2. coco-to-labelme: COCO JSON (or annotations dir) -> LabelMe sidecar JSON next to images.
3. coco-spec-to-yolo: COCO pose spec YAML -> YOLO pose dataset (data.yaml, images/, labels/).

Commands

LabelMe -> COCO keypoints

python annolid/core/agent/skills/dataset-convert/scripts/convert_dataset.py \
  labelme-to-coco \
  --input-dir /path/to/labelme_dataset \
  --output-dir /path/to/coco_out \
  --mode keypoints \
  --train-valid-split 0.8

COCO annotations dir -> LabelMe dataset

python annolid/core/agent/skills/dataset-convert/scripts/convert_dataset.py \
  coco-to-labelme \
  --annotations-dir /path/to/coco/annotations \
  --output-dir /path/to/labelme_out \
  --images-dir /path/to/coco/images

COCO pose spec -> YOLO pose dataset

python annolid/core/agent/skills/dataset-convert/scripts/convert_dataset.py \
  coco-spec-to-yolo \
  --spec-yaml /path/to/coco_pose_spec.yaml \
  --output-dir /path/to/yolo_pose_dataset

Notes

• labelme-to-coco keypoints mode requires pycocotools.
• For large datasets, prefer --link-mode hardlink (default) to avoid image copying.
• Always return output paths and summary counts after conversion.