💻 computer science

Badminton Technical Movement Evaluation: A New Benchmark and Baseline Model

This paper addresses the challenge of providing individualized feedback in large-scale university badminton classes by introducing a new benchmark dataset of 7,200 annotated video clips and a baseline model that automatically evaluates student movement quality against professional demonstrations, achieving promising performance metrics for intelligent physical education assistance.

Original authors: Kai Song, Songze Qiu, Ran Tong, Bing Ma, Nianchang Huang, Junxin Yang

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kai Song, Songze Qiu, Ran Tong, Bing Ma, Nianchang Huang, Junxin Yang

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

Imagine you're in a high school gym, and the badminton coach is trying to teach 67 students how to smash a shuttlecock. The coach is great, but they only have two hands and one set of eyes. They can't watch everyone at once, so some students are swinging their rackets like they're chopping wood, while others are actually hitting the bird. The coach is exhausted, and the students are guessing if they're doing it right.

Enter a new team of researchers who asked: "What if we could build a robot coach that watches every student, compares their swing to a pro's, and gives them a score?"

The Big Idea: A Digital Mirror
The researchers, led by Kai Song and Junxin Yang, didn't just dream this up; they built the tools to make it happen. They created a brand-new "benchmark," which is basically a giant, organized library of badminton videos. Think of it as a massive training camp where they filmed 5 professional coaches and 67 university students performing 10 different badminton moves (like the Forehand Clear or Backhand Drive).

Here's the cool part: they didn't use fancy, expensive motion-capture suits or studio cameras. They used smartphones. Why? Because that's what students actually have. They filmed in real places like dorm rooms and teaching buildings, not just perfect gyms, to make sure the system works in the messy real world. In total, they gathered 7,200 video clips.

How the "Robot Coach" Learns
The team built a computer model (a "baseline model") to act as the judge. You can think of this model as a super-observant referee with a split brain:

The Eye: It looks at a single frame of video to see what the player's body looks like right now.
The Memory: It connects the dots between frames to understand the flow of the movement, like how a dancer moves from one step to the next.
The Comparison: This is the magic sauce. The model takes the student's video and lines it up next to a "gold standard" video from one of the pro coaches. It doesn't just say "good job" or "bad job"; it measures the tiny differences between the student's swing and the pro's swing.
The Score: Finally, it translates those differences into a number on a 100-point scale.

What the Numbers Say (The Scoreboard)
The researchers tested their robot coach on a group of students it had never seen before. Here is how it performed, based on the numbers they reported:

The Average Mistake: On average, the robot's score was off by 7.47 points compared to the human coaches' scores. (Imagine if a human coach gave you an 85, the robot might have guessed 77 or 92).
The Agreement: The robot's ranking of who was doing better matched the human coaches 56% of the time in terms of order (SRCC of 0.56) and 58% in terms of linear correlation (PLCC of 0.58).
The "Close Enough" Wins: When the researchers checked how often the robot was within 10 points of the human score, it got it right 73% of the time (Acc10). That means for nearly three out of four students, the robot gave a grade that was very close to what a real teacher would give.

What It Got Right (and What It's Still Figuring Out)
The model was a whiz at judging moves like the Backhand Drive and Backhand Serve, where it scored very high on accuracy. However, it struggled a bit more with the Forehand Clear and Forehand Serve. The researchers suggest this might be because those specific moves are harder to capture or judge consistently, but they didn't say the system failed; they just noted it needs more work on those specific strokes.

The Verdict
The paper doesn't claim to have solved badminton forever or replaced human teachers. Instead, it suggests that this approach is feasible. It proves that you can use a smartphone and a computer program to give students a quick, personalized score on their technique, which could help teachers manage those huge classes better.

In short, the researchers built a digital mirror that can tell you, "Hey, your swing is a little different from the pro's, and here's a score to show you how close you are." It's not perfect yet, but it's a solid first step toward a future where every student gets a personal coach, even if that coach is made of code.

Technical Summary: Badminton Technical Movement Evaluation: A New Benchmark and Baseline Model

Problem Statement
University physical education in China faces a significant bottleneck in badminton instruction due to large class sizes and limited teacher resources. The traditional "teacher demonstration, student imitation" model struggles to provide personalized, one-on-one guidance for every student, hindering technical improvement and increasing instructor workload. While Artificial Intelligence (AI) has shown promise in video understanding and motion analysis, existing Action Quality Assessment (AQA) frameworks are primarily designed for professional sports with standardized scoring rules (e.g., gymnastics, diving) and single-video inputs. They lack the capability to perform fine-grained, exemplar-based comparisons between a student's performance and a coach's standard demonstration in real-world educational settings, particularly when using commodity devices like smartphones.

Methodology
To address these gaps, the authors propose a novel task: Badminton Technical Movement Evaluation, which aims to automatically score a student's movement quality by comparing it against standard coach demonstrations. The proposed solution consists of a new benchmark dataset and a baseline deep learning model.

Benchmark Dataset Construction:
- Data Collection: The dataset comprises 7,200 smartphone-captured video clips covering 10 fundamental badminton strokes (e.g., Forehand Clear, Backhand Drive, Net Drop).
- Participants: Data was collected from 5 professional coaches (providing standard exemplar videos) and 67 university students (performing the strokes in diverse real-world scenarios, including dormitories and teaching buildings, not just courts).
- Annotation: Expert coaches scored student performances on a 100-point scale using a multi-rater protocol (eliminating the highest and lowest scores to average the remainder).
- Split: The dataset is partitioned into a training set (53 students) and a test set (14 students) to ensure subject-independent evaluation.
Baseline Model Architecture:
The proposed model follows a four-stage pipeline to map a pair of videos (exemplar $V_E$ and student $V_C$ ) to a quality score $p$ :
- Intra-frame Feature Extraction: Utilizes a pre-trained Vision Transformer (ViT) to extract spatial discriminative features from each frame of both the coach's and student's videos, leveraging ViT's ability to capture long-range dependencies.
- Inter-frame Feature Fusion: Employs a Spatio-Temporal Fusion Module using a self-attention mechanism. This module concatenates frame-wise features and projects them into Query, Key, and Value spaces to capture temporal continuity and dynamics across the video sequence.
- Exemplar-Student Feature Comparison: Establishes an interactive relationship between the coach's and student's spatiotemporal features. This involves:
  - A subtraction operation to capture global discrepancies.
  - A bilinear pooling operation to capture local discrepancies.
  - Concatenation of these difference features followed by a linear transformation.
- Movement Scoring: A final linear layer predicts the quantitative quality score based on the fused comparison features.
- Training: The model is optimized using Root Mean Squared Error (RMSE) loss to minimize the deviation between predicted and ground-truth scores.

Key Contributions

Novel Task Definition: The paper formally defines the task of badminton technical movement evaluation as an exemplar-based comparison problem, distinct from standard single-video AQA tasks.
First Large-Scale Benchmark: The authors constructed the first specialized dataset for this task, featuring over 7,200 video sequences, 10 stroke categories, and professional manual annotations from over 60 students and 5 coaches, collected via mobile phones in uncontrolled environments.
Baseline Model: A novel deep learning architecture was proposed that integrates spatio-temporal representation learning with exemplar-based comparison to achieve automatic, quantitative scoring.

Experimental Results
The model was evaluated on the held-out test set using six metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Spearman's Rank Correlation Coefficient (SRCC), Pearson Linear Correlation Coefficient (PLCC), and accuracy within 5 and 10 points (Acc5, Acc10).

Overall Performance: The baseline achieved an MAE of 7.47 and an RMSE of 9.09. The correlation metrics were SRCC of 0.56 and PLCC of 0.58.
Accuracy: The model achieved an Acc5 of 38% and an Acc10 of 73%, indicating that the predicted scores fall within 10 points of the ground truth for the majority of samples.
Stroke-wise Analysis: Performance varied by stroke type. The model performed best on strokes like Backhand Drive (MAE 6.04, SRCC 0.75) and Backhand Serve (MAE 6.60). Performance was relatively lower on Forehand Clear and Forehand Serve, though positive correlations were observed across all strokes.

Significance and Claims
The paper claims that these results demonstrate the feasibility of automatic badminton skill assessment using AI. The proposed framework successfully bridges the gap between existing research and practical needs in AI-assisted physical education by enabling:

Automated, quantitative evaluation of student movements without requiring specialized sensors or controlled environments.
Reduction of manual workload for teachers in movement evaluation.
Provision of timely, targeted feedback to students, potentially improving the efficiency and quality of university badminton teaching.

The authors position this work as a foundational step, providing a reliable data foundation and a baseline model for future research into intelligent assistance systems for physical education.

Technical Summary: Badminton Technical Movement Evaluation: A New Benchmark and Baseline Model

More like this