FiftyOne / HuggingFace dataset bridge

[wkentaro]

What to build

A bridge between labelme's JSON and a dataset library (FiftyOne and/or HuggingFace datasets), so users can load labelme annotations into those ecosystems (and ideally back) without hand-rolling per-format exporters. Keep the plain COCO/YOLO file exporters for the file-oriented majority; this is an additional path for pipeline users.

This is HITL: scope (which library, which direction, dependency cost) is a maintainer decision.

Acceptance criteria

• labelme annotations can be loaded into at least one target (FiftyOne or HF datasets)
• The mapping (shapes, labels, masks) is documented
• The bridge is optional and does not add a hard dependency for core use

Blocked by

None - can start immediately