perp-cnn — Bowtip Detection

Detects the bow tip of boats in photofinish images. Used in perp_web for automated boat race timing.

Usage

import numpy as np
from perp_cnn import predict

image: np.ndarray = ...  # H×W×3, BGR or RGB
results = predict(image)

for box in results[0].boxes:
    x1, y1, x2, y2 = map(int, box.xyxy[0])
    conf = float(box.conf[0])

Or install as a Git dependency via Poetry:

perp-cnn = { git = "https://github.com/tillsc/perp_cnn.git" }

Model

• Architecture: YOLO11n
• Input: arbitrary image size (resized to 640px internally)
• Output: bounding boxes for class bowtip
• Training data: photofinish images from rowing races

Performance

Precision	100%
Recall	97.9%
Box accuracy (mAP50)	99.5%
Box precision (mAP50-95)	74.6%
CPU inference	~19ms / ~52 FPS

Training Notes

• No horizontal or vertical flips (boat orientation matters)
• No rotation or perspective warp
• Slight translation and scaling augmentation