dartscribe networks

The two networks of dartscribe. They find the bull and the crossings of a dartboard in a photo and tell where the 20 is. dartscribe uses them for auto calibration (RANSAC projection). Trained on geforcefan/dartscribe by Ercan Akyürek.

file	network	input	output
crossings.onnx	YOLO11n detect	1×3×960×960	boxes, 5 classes: bull, double outer, double inner, treble outer, treble inner
orientation.onnx	YOLO11n classify	1×3×512×512	20 classes: how many segments the 20 is turned from the top

Inputs

RGB, values divided by 255, NCHW, no NMS in the ONNX.

• crossings.onnx takes the photo letterboxed to 960×960 (gray padding 114). Output 1×9×18900: box center in rows 0 and 1, class scores in rows 4 to 8. The classes do not name the crossing, the board is symmetric. The projection has to come from a fit over all crossings.
• orientation.onnx takes the board warped into a 512×512 square with the projection, board center in the middle, 1.6 double outer radii to the edge. Output 1×20: class k means the 20 is k segments clockwise from the top.

Trained with tools/inference in the repository (Ultralytics, yolo11n.pt and yolo11n-cls.pt).