YOLOv5-RKNN2

Custom conversion in the Rockchip Model Zoo collection. Converted with RKNN Toolkit2 2.3.2. YOLOv5 n, s and m classifiers in FP16 and INT8 for RK3576 and RK3588. Download the model and metadata from the same directory under models/v2.3.2/classify/<variant>/<chip>/<precision>/.

The original classification weights came from the YOLOv5 v7.0 release and are included with their hashes and original URLs in source. The distribution includes the GPL-3.0 license from that release.

The custom conversion exports an ONNX graph and compiles it with RKNN Toolkit2. See source/convert-extended.py --help and use the v7.0 source checkout with --yolov5-source. INT8 uses the COCO128 calibration list in source/calibration.json. The manifest lists all output hashes.

The Model Zoo detection exports are currently unpublished while their upstream checkpoint provenance is being verified.

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

Model Zoo downloads

The models/v2.5.0/ downloads use RKNN Toolkit2 2.5.0. The conversion receipts identify the original ONNX and converter. See source and license notices.

Toolkit versions and tasks

• 2.3.2 has classification, detection, and segmentation.
• 2.5.0 has detection and segmentation.
• Detection includes n, s, m, and s-ReLU in FP16 and INT8 for both chips.

Toolkit 2.3.2 builds

See models/v2.3.2 for builds tested on their target board. Each conversion receipt identifies its source graph and compiler.