HYROX wall balls detection

Detects the medicine ball and the wall target in footage of HYROX wall balls. Built for a phone app that counts reps.

Not affiliated with or endorsed by HYROX.

detector-416.onnx on throws from four of our recordings, at their own speed. A ball is boxed in orange from a score of 0.3, a target in blue from 0.7. The four recordings are in the training data. A target the ball covers by half or more has no box, as in the training labels.

Labels

id name what it is
0 ball the medicine ball, in the hands or in the air
1 target the mark or plate on the wall the ball must hit

Files

The same model at three input sizes, and its weights. This is revision v10, trained on 14 November 2026.

file input use
detector-320.onnx 320 fastest, for a window around the target or tiles of the frame: the whole frame at 320 shows each object at half the size the model learnt
detector-416.onnx 416 faster, for every frame of a video
detector-640.onnx 640 more detail, for small or distant targets
detector.pt 640 the trained weights, to fine-tune or to export at another size with Ultralytics

Usage

With Ultralytics:

from ultralytics import YOLO

model = YOLO("detector-416.onnx", task="detect")
model.predict("frame.jpg", imgsz=416, conf=0.8, iou=0.1, save=True)  # writes runs/detect/predict/frame.jpg

Without it, with numpy, onnxruntime and opencv-python, python detect.py frame.jpg writes boxed.jpg:

import sys

import cv2
import numpy as np
import onnxruntime as ort

SIZE = 416  # 640 for detector-640.onnx
LABELS = ["ball", "target"]

image = cv2.imread(sys.argv[1])

# Shrink a picture wider than 640 pixels to 640 across, the width of the training frames.
if image.shape[1] > 640:
    image = cv2.resize(image, (640, round(image.shape[0] * 640 / image.shape[1])))

# Fit the picture into a grey square, without stretching it.
scale = SIZE / max(image.shape[:2])
h, w = round(image.shape[0] * scale), round(image.shape[1] * scale)
dx, dy = (SIZE - w) // 2, (SIZE - h) // 2
square = np.full((SIZE, SIZE, 3), 114, np.uint8)
square[dy : dy + h, dx : dx + w] = cv2.resize(image, (w, h))
pixels = square[:, :, ::-1].transpose(2, 0, 1)[None].astype(np.float32) / 255

# Each column is a candidate: centre x, centre y, width, height, then a score per label.
out = ort.InferenceSession(f"detector-{SIZE}.onnx").run(None, {"images": pixels})[0][0]
label, score = out[4:].argmax(0), out[4:].max(0)
boxes = np.stack([out[0] - out[2] / 2 - dx, out[1] - out[3] / 2 - dy, out[2], out[3]], 1) / scale

# Keep the candidates scoring 0.8 or more, and the best of those covering one object.
for i in cv2.dnn.NMSBoxesBatched(boxes, score, label, 0.8, 0.1):
    x, y, bw, bh = map(int, boxes[i].round())
    cv2.rectangle(image, (x, y), (x + bw, y + bh), (0, 200, 255), 2)
    cv2.putText(image, f"{LABELS[label[i]]} {score[i]:.2f}", (x, max(y - 5, 15)), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 200, 255), 2)
cv2.imwrite("boxed.jpg", image)

Training

Fine-tuned from Ultralytics' YOLO11n COCO weights, on wall-ball frames checked by hand and frames labelled with SAM 3. The training frames are not published.

Licence

AGPL-3.0, inherited from the Ultralytics YOLO11 weights.

Downloads last month
21
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for RodrigoQuelhas/hyrox-wallballs-detection

Finetuned
(214)
this model