💻 computer science

MonoClub: Pose-Guided Joint Localisation of Golf Clubhead and Shaft in Monocular Swing Video

This paper introduces MonoClub, a pose-guided framework that reliably localizes golf clubheads and shafts in monocular videos by integrating body-scale priors, geometric and appearance cues, and temporal consistency to overcome challenges like motion blur and occlusion.

Original authors: Zheng Zhang, Zhixin Zhou, Wen Jung Li, Rui Luo

Published 2026-10-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zheng Zhang, Zhixin Zhou, Wen Jung Li, Rui Luo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Watching a golf swing on a standard video camera is a deceptively difficult task for a computer. To the human eye, the motion is a fluid arc of power, but to a machine, it is a chaotic blur of thin lines and small, fast-moving objects. The golf club itself presents a unique puzzle: the shaft is a long, slender stick that often disappears into the background or blurs into a streak of motion, while the heavy clubhead at the end is a small, dark shape that can be easily confused with shadows or other objects. In the past, analyzing these swings required expensive sensors worn by the player or controlled studio environments with special cameras. The goal of recent computer vision research has been to achieve the same level of insight using only a single, ordinary video camera, but the visual noise of a real-world swing has made this a stubborn challenge.

A team of researchers has now developed a new system that solves this problem by treating the golfer's body and the golf club as a single, connected unit rather than two separate things to be found. Instead of trying to locate the club in isolation, their method first looks at the player's posture to understand where the hands are likely to be holding the club. This creates a starting point, or an anchor, that guides the search. The system then generates many possible guesses for where the club might be, testing each one against the visual evidence in the video frame. It checks if a proposed line for the shaft actually looks like a club shaft, if the head of the club matches the shape of a real club, and if the entire setup makes physical sense given the position of the golfer's hands. If the evidence is strong and consistent, the system accepts the observation; if the image is too blurry or the club is hidden, the system honestly admits it cannot see the club, rather than guessing and providing a wrong answer.

The researchers tested this approach on a collection of real golf videos taken from various angles and under different lighting conditions. They found that by linking the club's position to the golfer's body, the system could locate the clubhead with remarkable precision, missing the true center by an average of just 9.1 pixels on a standard video frame. In more than 93 percent of the cases where the club was visible, the system placed the clubhead within 10 pixels of its true location. It also proved highly effective at measuring the angle of the club shaft, estimating the direction with an error of only 3.6 degrees. Perhaps most importantly, the system learned to recognize when it was failing. In situations where the club was heavily blurred or blocked by the player's own body, the system achieved a visibility F1 score of 80.0%, indicating a strong balance between correctly identifying visible frames and avoiding false detections, rather than simply withholding judgment at a fixed rate.

This work represents a significant step forward because it moves away from trying to force a computer to see a thin object in a vacuum. Instead, it uses the context of the human body to make the search for the club much more reliable. The system does not need to understand the three-dimensional depth of the swing or reconstruct the physics of the ball flight; it simply needs to accurately describe the two-dimensional shape and position of the club as it appears on the screen. By combining the visual clues of the club's shape with the logical constraints of the golfer's grip, the researchers created a tool that is both accurate and honest about its limitations. The result is a method that can run quickly enough to be used in real-time applications, offering a robust way to analyze golf swings from ordinary video without the need for specialized equipment or sensors.

Technical Summary: MonoClub

Problem Statement

Reliable observation of a golf club in ordinary monocular video is inherently difficult due to the physical characteristics of the object: the shaft is a long, thin structure prone to motion blur and self-occlusion, while the clubhead is small and often blurred. Traditional approaches relying on independent object detection or keypoint regression are brittle; a visually salient line may belong to the background, and a plausible endpoint may be paired with an incorrect shaft. Furthermore, existing datasets and methods often lack dense geometric annotations or treat the club as an isolated object rather than a configuration coupled with the player's body. The core challenge addressed by this work is how to couple clubhead and shaft evidence within a single, unconstrained golf swing video to produce a coherent two-dimensional configuration.

Methodology

The paper proposes MonoClub, a pose-guided joint localisation framework that estimates a coherent 2D club configuration (grip anchor, shaft centreline, and clubhead location) from a single video frame. The system operates on a "propose-and-rank" workflow rather than a single end-to-end prediction.

1. Pose-Conditioned Grip Anchor

The system first estimates the player's body pose and hand locations (using YOLOv8-Pose and optionally MediaPipe). This provides a soft grip anchor ( $g_t$ ) and a body-scale factor ( $b_t$ ). This anchor is not a fixed 3D reconstruction but a 2D prior that constrains where the shaft is likely to originate, preventing arbitrary background lines from being accepted as the club origin.

2. Candidate Generation

Multiple candidate club configurations are generated. These hypotheses consist of a grip location, a shaft centreline segment, and a clubhead endpoint. Candidates are proposed using:

Prior states from previous frames.
Classical geometric cues (LSD line segments, Hough transforms, dark connected components).
Radial rays emitted from hand pixels.
Learned clubhead proposals.

3. Joint Evidence Scoring

Instead of selecting the best local detection, the system scores competing hypotheses using a unified function ( $S_{t,k}$ ) that integrates:

Shaft-Strip Evidence: A pose-conditioned TinyUNet predicts a probability map for the shaft along the candidate line. A "soft-clDice" loss is used during training to ensure topological continuity of the thin shaft.
Clubhead Endpoint Evidence: A separate TinyUNet predicts the clubhead mask near the distal endpoint of the candidate shaft.
Geometric Compatibility: Terms penalize candidates that are physically implausible relative to the body scale or the grip anchor (e.g., detached shafts).
Temporal Consistency: An optional term (via beam-search dynamic programming) discourages abrupt, unsupported switches in state between frames, stabilizing predictions during fast motion.

4. Reliability Gating and Abstention

A critical component is the reliability gate ( $q_t$ ). The system does not force a prediction when visual evidence is ambiguous. It calculates a reliability score based on geometric residuals, temporal inconsistency, and candidate disagreement.

Accepted: High candidate score and high reliability.
Observation-Only: Raw candidates are shown, but no derived conclusions are made.
Abstain: The system explicitly reports no measurement for frames with motion blur, self-occlusion, or out-of-frame clubs. This allows for the evaluation of "visibility" rather than silently discarding difficult frames.

Key Contributions

Grip-Anchored Joint Formulation: The paper treats the clubhead, shaft, and visibility as a single, jointly estimated configuration rather than independent outputs. This enables consistency checks between the grip region, the shaft line, and the endpoint.
Candidate-Guided Inference Workflow: The method integrates pose context, learned shaft-strip segmentation, and geometric compatibility within a unified scoring function. Classical cues are used only for proposal generation, while deep learning provides the primary evidence for ranking.
Source-Video-Disjoint Evaluation Protocol: The evaluation is performed on a dataset split at the video level (no adjacent frames from the same video appear in different splits). This prevents temporal leakage and provides a realistic assessment of generalization. The protocol reports proposal recall, segmentation quality, end-to-end localisation, and visibility F1 scores with 95% confidence intervals.

Experimental Results

The framework was evaluated on a custom dataset of 60 monocular golf swing videos (20 for testing, 40 for training/validation) and the public GolfClub dataset.

End-to-End Localisation: MonoClub achieved a Mean Head Error (MHE) of 9.1 pixels and a PCK@10 of 93.5%, outperforming state-of-the-art baselines like CADDIE (MHE 11.2 px) and RacketVision.
Shaft Geometry: The method achieved a shaft angle error of 3.6° and an endpoint-to-centreline distance of 6.7 pixels.
Segmentation: The clubhead segmentation module achieved a Dice coefficient of 0.84 and IoU of 0.73.
Candidate Recall: The proposal generation step achieved high recall (96.2% at K=5), ensuring the true clubhead is rarely missed before the scoring stage.
Efficiency: The model uses 12.4 million parameters and runs at 52 FPS on an RTX 4090, making it suitable for real-time applications.
Ablation Studies: Removing pose conditioning increased MHE by 68%, and removing joint coupling significantly degraded shaft angle estimation, confirming the necessity of the coupled approach.

Significance and Claims

The paper claims to provide a reproducible basis for robust two-dimensional golf-club observation. Its significance lies in:

Robustness to Ambiguity: By coupling the shaft and clubhead with body pose and employing a reliability gate, the system handles motion blur, self-occlusion, and background clutter better than independent detectors.
Explicit Uncertainty Handling: The ability to abstain from making predictions when confidence is low (rather than outputting noisy data) is a key feature for downstream applications like coaching or visualization.
Downstream Utility: While the primary contribution is 2D observation, the framework supports downstream tasks such as swing visualization (e.g., overlaying professional references).

Limitations and Scope:
The authors explicitly state that this is a two-dimensional method. It does not claim to perform monocular depth reconstruction, 3D biomechanical analysis, or professional swing reprojection. These are identified as downstream applications. The method relies on accurate 2D body pose estimation and assumes a rigid club model; it may struggle with severely bent shafts or non-standard grips. The evaluation is strictly limited to 2D metrics, and the professional-reference overlays shown are qualitative visualizations, not validated 3D reconstructions.