This paper presents a robust, interpretable pipeline for detecting specific golf swing faults from ordinary smartphone video. By eschewing "black-box" end-to-end learning in favor of biomechanically-grounded features and shallow boosted trees, the authors achieve high accuracy (up to 0.975 ROC-AUC) while maintaining the explainability required for coaching and injury prevention.
TL;DR
The authors developed an AI pipeline that identifies common golfing faults—"Over-the-Top" and "Early Extension"—using standard monocular video without markers. By tracking body pose and clubhead trajectories to calculate coach-validated biomechanical features, the system achieves high detection accuracy and runs in seconds on consumer hardware.
What Problem Does This Paper Address?
In golf, technique is the difference between a high-performance shot and a chronic back injury. While professional laboratories use expensive marker-based motion capture to analyze swings, most golfers rely on a coach’s subjective eye or their own smartphone recordings.
Existing AI approaches for sports analysis often provide a generic "quality score" or focus on sequencing (e.g., identifying the moment of impact). However, they rarely detect specific, named technical faults. The challenge lies in the fact that these faults—like the clubhead deviating from the "swing plane"—are dynamic, biomechanical constructs that are difficult for standard pose estimators to capture directly from 2D video.
Methods
The researchers proposed a pipeline that prioritizes interpretability over raw complexity:
- Detection & Tracking: They use MediaPipe Pose for body landmarks and a fine-tuned YOLO11-L detector for the clubhead. A gated Kalman filter and RTS smoother bridge detection gaps (e.g., when the club moves too fast for the shutter).
- Normalization: To handle different camera distances and golfer sizes, coordinates are mapped to a body-centered, aspect-corrected frame based on the golfer's height and hip position.
- Phase Alignment: The downswing is resampled into 30 discrete "phase steps" from the top of the backswing to impact. This allows the model to compare swings regardless of their absolute tempo.
- Feature Engineering: Instead of feeding raw pixels to a neural network, they calculate 11–15 biomechanical signals per phase step (e.g., clubhead height above the swing plane, spine angle, shoulder tilt).
- Classification: These trajectories are fed into XGBoost (shallow boosted trees). This choice is crucial: it allows researchers to generate "importance maps" to see exactly which signal at which moment of the swing triggered a "fault" detection.
Datasets
The study utilized a combination of professional and amateur data:
- OTT Dataset: 160 clips. Positives (faulty) were sourced from Reddit and YouTube; negatives (clean) were largely professional swings from the GolfDB dataset, supplemented by amateur clips from TikTok.
- EE Dataset: 143 clips. A similar mix of amateur "faulty" clips and professional "clean" clips.
- Clubhead Training: The YOLO11 detector was fine-tuned on 700 hand-labeled frames to distinguish the ball, handle, and clubhead.
Key Insights and Ideas
- The "Smoking Gun" Feature: For the Over-the-Top (OTT) fault, the single most informative signal was the clubhead's height above the swing plane (AUC 0.940). This aligns perfectly with how human coaches diagnose the fault.
- Timing is Everything: Faults are not visible throughout the entire swing. The model's importance maps show that OTT evidence is concentrated between 65% and 80% of the downswing, while Early Extension (EE) evidence peaks in the final 20% before impact.
- Avoiding "Shortcut Learning": The authors compared their model to a deep 3D-CNN (R3D-18). While the CNN was accurate, it "cheated" by identifying faults from the golfer's address posture or environment before the swing even started. The biomechanical model, by contrast, fell to chance levels when movement was removed, proving it actually learns the mechanics of the fault.
Key Results
- High Accuracy: The pipeline detected OTT with a cross-validated ROC-AUC of 0.975 and EE with 0.911.
- Explainability: Because the features are biomechanical, the model can tell a golfer why they were flagged (e.g., "Your clubhead was 0.3 body heights above the plane at mid-downswing").
- Efficiency: Running on an Apple M2 laptop, the system processes a swing in roughly 13 seconds. Using smaller YOLO variants (YOLO11-N) could bring this down to near-instantaneous feedback (56 fps).
Limitations and Future Work
- Skill Confound: Because most "clean" swings were professionals and "faulty" swings were amateurs, the model might partially be learning to distinguish pros from amateurs rather than just the fault itself.
- Label Noise: The clips were not graded by a panel of PGA professionals; labels were inferred from the context of the source (e.g., an amateur asking for help with a specific slice).
- Single Viewpoint: The system currently only works for "down-the-line" camera angles.
- Future Directions: The authors suggest expanding the dataset to include more "clean" amateur swings and using 3D body reconstruction to capture rotations that are currently hidden in 2D video.
Summary updated