Aug 16, 2026 · 3 min

MyShot

Smartphone video to 3D swing and shot outcome · clubhead-speed median error 6.1%

Object Detection · Human Pose · 3D Vision · Ball-flight Prediction

Built with YOLO11 · RTMPose · MotionAGFormer · Physics-based ball flight · Reliability scoring

Frame-aligned AIHub research-data example · face anonymized · golfer, club, ball, pose, and angle tracking

WHY

Golf coaching and equipment are expensive, and practicing alone makes it difficult to see both what is wrong with a swing and how that movement changes ball speed, direction, and carry. Most pose tools stop at drawing body motion. I wanted one smartphone video to connect swing correction with the outcome of the shot.

Clubhead speed 6.1%

Median error on GolfPose Vicon motion

3D reconstruction 35.6 mm

MPJPE · X-Factor MAE 3.1°

Carry ground truth 25 pairs

Built from 50 AIHub swing clips

HOW

From one phone video to posture, speed, direction, and carry

The system treats body motion, club motion, and the ball as one sequence, then separates measured signals from calibrated and physics-based estimates.

01 Golfer, club, and ball
Track the three targets through address, top, impact, and finish instead of analyzing the golfer alone.
02 2D to 3D motion
Lift RTMPose joints with a temporal MotionAGFormer model trained with GolfPose Vicon and pseudo-3D swing data.
03 Swing mechanics
Compute joint movement, X-Factor, swing power, club path, and clubhead speed across the full motion.
04 Personal calibration
Use physique and per-club carry or ball-speed records as calibration inputs rather than forcing one population model on every golfer.
05 Shot outcome and reliability
Estimate ball speed, start direction, and carry with a physics-based ball-flight model, while pose disagreement and jitter flag unreliable frames.

RESULT

An anonymized four-phase golf swing example showing the source frame, detected 2D joints, and reconstructed 3D skeleton at address, top, impact, and finish — Single-camera input · 2D joints · current-model 3D reconstruction

Six small-multiple plots comparing predicted and ground-truth normalized X-Factor curves on independent CMU golf motion-capture trials — Independent CMU motion capture · six trials · mean absolute correlation 0.95

Contribution

Built the path from target tracking and 2D-to-3D motion to joint metrics, clubhead speed, personal calibration, and a physics-based ball-flight model.
Found that body-only 30 fps motion was too weak for direct carry regression, then redesigned the system around ball and club tracking plus per-club calibration.

Evidence

Private project