ELDET

Reference implementation of ELDET (Early-Learning Distillation with Noisy Labels for Object Detection) for YOLO and RF-DETR. The shared logic lives in eldet/; trainers wrap the vendored stacks in ultralytics/ and rf-detr/.

Install

Python 3.10 or newer. From the repository root:

pip install -e .
pip install -e "./rf-detr[train]"
pip install -e "./ultralytics"

Train

YOLO (downloads the small coco8 sample if needed):

python examples/train_yolo_eldet.py --data coco8.yaml --epochs 3

RF-DETR (pass a COCO-style dataset root with train/, valid/, and annotations):

python examples/train_rfdetr_eldet.py --dataset-dir /path/to/coco --dry-run
python examples/train_rfdetr_eldet.py --dataset-dir /path/to/coco --epochs 10

Optional: pytest eldet/tests -q after pip install -e ".[dev]".

License

First-party code in eldet/, examples/, and configs/ is MIT. Vendored Ultralytics is AGPL-3.0; vendored RF-DETR is Apache-2.0. See VENDOR_VERSIONS.txt for pinned revisions.