---

name: dataset-annotation description: "Pseudo-label image folders with GroundingDINO detections, optional COCO export, and annotated-image review while keeping optional dataset tooling explicit." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Dataset Annotation

Use this sub-skill when the task is to pseudo-label a folder of images with GroundingDINO, inspect annotated review images, or export detections as a COCO-style dataset.

Route first

• For one image, direct API inference, token spans, or prompt debugging, use ../inference/.
• For COCO AP or benchmark evaluation, use ../evaluation/.
• For Gradio, notebooks, Grounded-SAM, or image-editing integrations, use ../integrations/.

What this covers

• Folder pseudo-labeling with image_directory, text_prompt, box_threshold, text_threshold, weights_path, config_path, and optional subsample controls.
• Optional COCO export and annotated-image review outputs without launching a GUI by default.
• Safe handling for optional fiftyone and Typer-style workflow dependencies; the bundled helper uses argparse, not Typer.
• Conversion from GroundingDINO normalized cxcywh predictions into FiftyOne relative xywh detections.

Bundled helper

From this sub-skill directory, start with:

python scripts/grounding_dino_pseudolabel.py --help

Minimal non-GUI COCO export example:

python scripts/grounding_dino_pseudolabel.py \
  --image-directory ./images \
  --text-prompt "bus . car ." \
  --config-path ./GroundingDINO_SwinT_OGC.py \
  --weights-path ./groundingdino_swint_ogc.pth \
  --export-coco \
  --draw-labels \
  --output-dir ./pseudolabel-output \
  --subsample 25

The helper never downloads checkpoints or datasets. Provide an existing config file and checkpoint, and pass --view only when a local FiftyOne GUI/server is intended.

References

• Workflow and commands: references/pseudolabel-workflow.md
• Detection and export formats: references/data-formats.md
• Failure modes and recovery: references/troubleshooting.md