YOLOv10-RKNN2-Ultralytics

Ultralytics export for RK3576 and RK3588. YOLOv10 n, s, m, b, l, and x in FP16 and INT8 for RK3576 and RK3588. Choose a model under models/v2.3.2/detect/<variant>/<chip>/<precision>/ and download both model.rknn and metadata.yaml.

These files were converted from the official Ultralytics assets. The Ultralytics exporter and RKNN Toolkit2 changed the model format and, for INT8, quantized the graph. The RKNN exports and original checkpoints are distributed under AGPL-3.0. See LICENSE and checkpoint origins.

Reproduce the conversion

The exact input checkpoints, conversion scripts, package versions, and calibration hashes are in source. The Ultralytics 8.4.184 source distribution is available on PyPI. The converter uses RKNN Toolkit2 2.3.2, available separately from Rockchip.

Run python source/convert-model.py source/weights/yolov10n.pt --chip rk3588 --precision fp16 --imgsz 640 --output yolov10n-rk3588. For INT8, use --precision int8 --data /path/to/coco128.yaml --calibration-limit 128. The calibration manifest links to the source images. Point the YAML at the extracted COCO128 tree and its train2017 image directory.

The model manifest records file hashes, precision, and input size.

INT8 detection

Use the metadata.yaml from the same INT8 model folder. The detector outputs normalized box coordinates, so the app needs coordinates: normalized. Import both files together in the warehouse. For the CLI, keep them in the same directory. Interpreting the coordinates as pixels can result in no boxes.

Models are organized by RKNN Toolkit2 version under models/v<version>/. Each model and its metadata share a folder.