Real-time golf club head detection

Smaller than a hand, faster than 150 km/h. We find it in every frame.

The club head is the hardest object in golf: small, metallic, motion-blurred and faster than 150 km/h at impact. For DeepSwing we trained our own detection model that reliably finds it in high-speed footage. It runs as a Core ML model directly on the iPhone's Neural Engine.

Fig. 01 — Club head detection
high-speed capture
240fps
exposure against motion blur
1/8000s
needed on the club
0sensors

01.1 — Object detection

Club head detection

01

Trained on what cameras really see

Club heads look different in every video: driver, iron, putter. Sun, indoor, artificial light. From the front and from behind. Our dataset covers that variety. The model is built on a current YOLO architecture, converted and optimised for Apple's Neural Engine.

02

From points to a path

The positions per frame become the trail: the path of the club head through the whole swing. From it we read swing plane, club path and tempo and lay the result over the video. Every point can be checked in the editor.

03

Milliseconds before impact

In the launch monitor, detection measures club head speed in the last frames before impact. Combined with ball speed this gives the smash factor: how efficiently the ball was struck.

01.2 — Capabilities

What it does.

In use in our products, every day, on real devices.

  1. 01

    No sensor, no markers

    The club head is found in the normal camera image. No sticker, no clip on the shaft.

  2. 02

    Tracked frame by frame

    From address to finish, even with motion blur and backlight.

  3. 03

    Club speed

    The club head's speed just before impact, as a value in every launch monitor shot.

  4. 04

    Trail through the whole swing

    The club head's path as a line over the video, checkable point by point in the editor.

  5. 05

    Swing plane

    The plane in the backswing and downswing, with a hint when the club comes too steep or too flat.

  6. 06

    "Stuck" detected

    When the club comes from too far inside, the coach explains exactly that and shows the right drills.

  7. 07

    Compare and export

    Two trails side by side, overlay videos to share with a coach or the community.

  8. 08

    Every club

    Driver, woods, irons and wedges, indoors and outdoors.

01.3 — How it works

How it works

  1. 01

    Capture

    240 frames per second, 1/8000 s exposure, 1080p.

  2. 02

    Detection

    YOLO model on the Neural Engine, frame by frame.

  3. 03

    Tracking

    Association over time, outliers are rejected.

  4. 04

    Measurement

    Path, plane and speed of the club head.

01.5 — FAQ

Frequently asked questions

Does club head detection need markers or sensors?

No. The model detects the club head in the normal camera image. No stickers, no sensors on the club and no extra hardware.

Does detection run offline?

Yes. The model runs on the device. Videos do not leave the phone for analysis.

Can you build such models for other sports equipment?

Yes. Clubs, rackets, balls, pucks and tools follow the same pattern: capture data, label, train, optimise for the target device and measure against a reference.

Contact

What should your camera recognise?

Object detection, body tracking, segmentation or cartography, on your device, in real time. Let's talk about it.