🤖 AI

BadmintonGRF: A Multimodal Dataset and Benchmark for Markerless Ground Reaction Force Estimation in Badminton

This paper introduces BadmintonGRF, a multimodal dataset and benchmark featuring synchronized high-frame-rate multi-view video and instrumented ground reaction force data from 10 subjects, designed to advance markerless load estimation in badminton through curated impact segments, time-aligned metadata, and comprehensive baseline evaluations.

Original authors: Kuoye Niu, Jianwei Li, Shengze Cai, Yong Ma, Mengyao Jia, Lishun Shen, Zhenheng Zhang, Yuxin Peng, Xian Song

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kuoye Niu, Jianwei Li, Shengze Cai, Yong Ma, Mengyao Jia, Lishun Shen, Zhenheng Zhang, Yuxin Peng, Xian Song

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to teach a computer to understand exactly how hard a badminton player is hitting the floor every time they jump and land. Usually, to get this information, you need the player to stand on special, expensive, laboratory-grade scales (called force plates) while wearing sensors. This is great for a science lab, but it's impossible to do during a real, fast-paced game or a regular training session.

This paper introduces BadmintonGRF, a new "training manual" for computers that solves this problem. It's a massive collection of data that teaches AI how to guess those floor forces just by watching a video, without needing any sensors on the player's body.

Here is a breakdown of what they did, using some everyday analogies:

1. The "Time-Travel" Puzzle (Synchronization)

In a real lab, the video cameras and the floor scales usually run on the exact same clock. But in this project, the researchers used eight regular, high-speed consumer cameras (like the ones on your phone or a GoPro) alongside the lab scales.

The Problem: These cameras don't share a clock. One might be a split-second faster than the other. It's like trying to conduct an orchestra where the drummer and the violinist are listening to different metronomes.
The Solution: The team built a system to manually and automatically line up the video and the force data. They treated it like a puzzle, finding specific moments (like a foot hitting the ground) to sync the two streams perfectly. They even added a "confidence score" to tell users how sure they are about the timing.

2. The "Two-Layer" Library (The Dataset)

To protect the athletes' privacy while still helping scientists, the data is split into two "tiers," like a library with a public reading room and a restricted archive:

Tier 1 (The Public Reading Room): This is what everyone can download for free. It contains the "skeleton" of the players (where their joints are in the video) and the ground force numbers. It's like giving someone a stick-figure animation and the data, but not the actual video footage. This is enough to train AI models to predict forces.
Tier 2 (The Restricted Archive): This contains the raw, high-definition video and the full 3D motion capture data. You can only get this if you apply for permission. This is for researchers who need to study the players' actual appearance or clothing, not just their movement.

3. The "Stress Test" (The Benchmark)

The researchers didn't just dump the data; they created a strict test to see how well different AI models perform.

The Challenge: They used a "Leave-One-Subject-Out" method. Imagine you teach a student badminton with nine different players, and then you test them on the tenth player they've never seen before. This tests if the AI has truly learned the physics of badminton, or if it just memorized the specific movements of one person.
The Results: They tested 10 different AI "brains" (models). The best ones could predict the vertical force with about 40% accuracy (measured by a statistical score called $r^2$ ). While this isn't perfect, it's a solid start for a sport as chaotic and fast as badminton.

4. Why Badminton?

Most previous studies looked at walking or jumping in a straight line (like a treadmill). Badminton is different; it's a "stop-and-go" sport with sudden stops, sharp turns, and heavy landings.

The Analogy: If previous datasets were like a recipe for baking a simple loaf of bread, BadmintonGRF is like a recipe for a complex, multi-layered cake that requires precise timing and temperature changes. It captures the messy, real-world stress on a player's legs.

What This Actually Means (Based strictly on the paper)

It's a Resource, Not a Magic Wand: The paper emphasizes that this is a tool for researchers. It provides the data, the code to load it, and the rules for testing.
It Solves a Specific Gap: Before this, there was no public dataset that combined high-speed video, lab-grade force measurements, and badminton-specific movements in a way that researchers could easily compare their results.
It's About "Markerless" Estimation: The goal is to eventually let coaches or players use just a video camera to estimate how much stress their legs are taking, without needing a $50,000 lab setup.

In short, BadmintonGRF is a carefully organized, high-quality "training ground" that allows computers to learn the relationship between a badminton player's video movements and the invisible forces hitting the floor, all while keeping the athletes' identities safe.

Technical Summary: BadmintonGRF

Problem Statement

Accurate estimation of Ground Reaction Forces (GRF) is critical for monitoring workload, analyzing fatigue, and assessing lower-extremity injury risks in sports. While markerless estimation (deriving GRF from video or pose) is advancing, it currently lacks robust validation in non-periodic, high-intensity court sports like badminton. Existing resources predominantly focus on periodic gait or scripted jumps, failing to capture the irregular, high-impact footwork characteristic of competitive play. Furthermore, there is a scarcity of public datasets that pair instrumented, laboratory-grade GRF with high-frame-rate, multi-view video, particularly where the video and force sensors are not hardware-synchronized (genlocked). This gap limits the development of markerless load estimation models for realistic training environments.

Methodology and Dataset Construction

The BadmintonGRF dataset addresses these gaps through a study-grade multimodal capture pipeline involving 17 competitive badminton athletes (national second-tier or higher).

Data Acquisition

Sensors: The setup includes eight fixed RGB cameras (DJI Osmo Action 4, ~120 FPS), four Kistler six-axis force plates, and an eight-camera Vicon motion capture system (exported to C3D).
Synchronization Challenge: Unlike traditional lab setups, the consumer RGB cameras operate on independent oscillators without hardware genlock to the laboratory time base.
Alignment Solution: The authors developed a software-based alignment pipeline. This involves:
1. Human-Verified Events: Annotators pair salient GRF peaks with video events to estimate per-camera time offsets.
2. Automated QA: A pipeline screens for offset consistency and flags anomalies (5.2% of trials required manual reconciliation).
3. Uncertainty Metadata: Released metadata includes offsets and derived uncertainty values.
Protocol: Athletes performed structured drills and match-like rallies under both fresh and induced-fatigue conditions. The protocol cycles through rally play and specific footwork modules (upper-net, smash-oriented, six-point reactions).

Data Processing and Segmentation

Ground Truth: Reference GRF is aggregated from the four plates, with vertical force ( $F_z$ ) serving as the primary supervision target due to its dominance in injury surrogacy.
Segmentation: The pipeline detects landing peaks in the aligned $F_z$ signal and crops symmetric windows of $\pm0.5$ seconds (approx. 120 frames).
Quality Gates: A reference loader filters instances based on:
- Lower-body mean pose score $\ge 0.70$ .
- Tracking loss rate $\le 0.05$ .
- Peak $F_z$ within three times body weight.
- Low-score masking and finite-difference sanitization.

Benchmark Design

The dataset is released in two tiers to balance utility and privacy:

Tier 1 (Public): Contains time-aligned pose data, processed GRF, metadata, and predefined splits under CC BY-NC 4.0. This tier supports the primary benchmark without requiring raw video or C3D files.
Tier 2 (Controlled Access): Provides raw RGB and C3D data for studies requiring appearance analysis or full kinematics.

The benchmark employs a Leave-One-Subject-Out (LOSO) protocol to evaluate cross-subject generalization, a critical metric for real-world applicability.

Key Contributions

Study-Grade Multimodal Capture: The first public release, to the authors' knowledge, combining eight-view ~120 FPS RGB, four Kistler force plates, and Vicon C3D for non-periodic badminton impacts, complete with fatigue-aware protocol tags.
Software Alignment without Genlock: A robust "human-in-the-loop" pipeline achieving roughly frame-level alignment between consumer RGB and laboratory sensing, complete with automated QA and uncertainty metadata.
Packaged Benchmark: A curated release of 17,425 impact segments (12,867 after quality gates; 1,732 unique impacts) with deterministic splits and loaders.
Transparent Evaluation: A fixed LOSO protocol, four complementary metrics ( $r^2$ , RMSE, peak magnitude error, peak timing error), and ten documented reference baselines to minimize implementation variance.

Results and Benchmark Performance

The paper evaluates ten reference architectures (including Temporal Convolutional Networks, Transformers, and Graph Neural Networks) on the Tier 1 benchmark.

Performance Ceiling: The best-performing models (PatchTST and ST-GCN+Transformer) achieve an $r^2$ of approximately 0.40 in the LOSO single-view setting.
Dominant Factors: Analysis indicates that cross-subject biomechanical variability is the primary source of error, dominating over residual synchronization noise. Synthetic timing stress tests showed that a 1-frame shift reduces $r^2$ by ~0.05, whereas held-out subjects induce swings >0.15.
Metric Trade-offs:
- PatchTST leads in curve fitting ( $r^2$ ) and RMSE.
- ST-GCN+Transformer leads in peak magnitude and timing accuracy, likely due to its explicit graph structure handling short landing transients.
- Late Fusion: Combining multi-view predictions can improve $r^2$ (e.g., TSMixer reaching 0.472) but often degrades peak metrics if per-view predictions conflict.
Within-Trial Performance: When subjects are seen during training (Within-trial setting), $r^2$ improves to ~0.60, confirming that the current LOSO ceiling is driven by subject transfer challenges rather than alignment issues.

Significance and Claims

The authors position BadmintonGRF as a necessary resource to bridge the gap between controlled laboratory biomechanics and unconstrained sports training.

Reproducibility: The paper emphasizes that the resource's value lies in its "audited" nature—providing explicit alignment protocols, quality assurance scripts, and deterministic loaders that allow researchers to reproduce results without hidden preprocessing choices.
Multimedia Relevance: Beyond sports science, the dataset serves as a stress test for multimedia research regarding heterogeneous clocks, occluded pose-to-dynamics modeling, and domain shifts between lab and training venues.
Modest Claims: The authors explicitly state that the reference models serve as "reproducibility anchors" rather than claims of state-of-the-art performance. They acknowledge that the current performance ceiling is limited by subject variability and that future gains will likely require subject-aware or metadata-conditioned objectives rather than further alignment tweaks.
Accessibility: By providing a Tier 1 dataset with privacy-preserving pose data, the authors aim to lower the entry barrier for researchers interested in pose-to-force estimation without needing access to restricted raw video or expensive laboratory setups.