MotionCaddie Try the demo
MIDS W210 Capstone · UC Berkeley School of Information

See your golf swing in 3D
from one phone video.

MotionCaddie reconstructs a full 3D swing from a single 2D clip, measures it against tour-pro reference ranges, tracks the ball's actual flight when it's visible in frame, and explains what it found in plain English — grounded only in what was actually measured, never invented advice.

Swing feedback that needs data shouldn't need a lab

Objective, biomechanical feedback on a golf swing — not just "your shoulders open early" but an actual number for it — has always needed equipment most golfers never get access to.

What it takes today

Multi-camera launch monitors, marker suits, or an in-person session with a biomechanics coach. A single phone video, on its own, has no depth — so the 3D angles that actually matter (hip-shoulder separation, posture, weight shift) can't be measured from it directly.

What golfers actually have

A phone. That's it. Amateur players and coaches who want objective, repeatable feedback need a way to get real 3D biomechanical measurement out of ordinary 2D video — without buying anything or booking a lab session.

What MotionCaddie does

Upload one clip. Get a 3D replay, a tour-referenced scorecard, and a coach that only talks about what was measured.

3D from a single camera

A 2D pose model finds every joint in the frame, then a 3D lifting model reconstructs the depth a flat video can't show — no extra hardware.

Tour-referenced indicators

Tempo, X-factor, posture, weight shift and more, computed at each swing event and scored against a tour-pro reference distribution.

Measured ball flight, not a guess

When the ball is visible in frame, a subpixel tracker follows it and fits a physics model to the real trajectory — reporting speed, launch angle, and carry with a confidence band. Can't see the ball? It falls back to a club-based estimate, clearly labeled as such.

A coach that won't make things up

The explanation layer is only allowed to describe measured numbers. A grounding verifier checks every claim before it reaches the golfer.

How it works

One phone video goes through six stages to become a coaching scorecard — while a parallel stage tracks the ball itself.

1

2D pose

MediaPipe locates every joint, every frame.

2

Identity repair

A Viterbi pass fixes left/right limb swaps before anything is lifted to 3D.

3

3D lift

A pose-lifting model reconstructs the depth a single camera can't see.

4

Event detection

A trained model finds address, top, impact and follow-through.

5

Indicators

15 biomechanical measurements, scored against tour reference.

6

Grounded explanation

An LLM writes a plain-English read, verified against the numbers.

Read the full technical deep dive — model architectures, benchmark numbers, the ball-tracking physics fit, and the grounding verifier →

Where the accuracy comes from

The event detector — the model that finds address, top, impact and follow-through — was the project's biggest technical win.

0.865
PCE@5 on held-out swing-event detection — beats the published GolfDB SwingNet baseline (~0.76)
15
Biomechanical indicators computed per swing and scored against tour-pro reference ranges
0%
Invented advice — every coaching claim is checked against the measured numbers before it's shown

Evaluated on 1,400 labeled clips from GolfDB across ~28,000 model × clip runs comparing pose backbones, 3D lifters, and event detectors. Full methodology and the complete leaderboard are in the repo's RESULTS.md.

Team

MIDS W210 Capstone Project, UC Berkeley School of Information — 2026

Austin Lee

Banjot Saini

Theodore Hui

Lawrence Jordan

Ready to see your swing measured?

Upload a clip or open one of the pre-analyzed demo swings.

Try the live demo →