asadahsan148/scos-corner-detector

SCOS Snooker Table Corner Detector

Detects the 4 corners of a snooker table (TL, TR, BR, BL) for automatic perspective calibration in the SCOS (Snooker Club Operating System).

Replaces manual click-to-calibrate with a single model inference call.

Model Details

• Architecture: YOLOv8s-pose (keypoint detection)
• Keypoints: 4 (Top-Left, Top-Right, Bottom-Right, Bottom-Left)
• Image size: 640x640
• Class: table (1 class)

Usage

from ultralytics import YOLO

model = YOLO('asadahsan148/scos-corner-detector')  # auto-download from HF Hub
results = model('frame.jpg')

# Get corners: [TL, TR, BR, BL]
corners = results[0].keypoints.xy[0].cpu().numpy()
# corners[0] = TL, corners[1] = TR, corners[2] = BR, corners[3] = BL

SCOS Integration

The SCOS backend auto-calibration route calls the HF Space inference endpoint, which runs this model and returns the 4 corner coordinates directly. These are fed into the existing perspective warp pipeline without any manual input.

Training Data

Annotated frames extracted from live CCTV footage of snooker tables. Labels: 4 keypoints per frame in YOLO pose format.