ShuttleSet: Badminton Stroke-Level Analytics

Updated 9 July 2026

ShuttleSet is a comprehensive badminton singles dataset with 36,492 strokes annotated at a stroke level, capturing tactical, spatial, and temporal details.
It enables turn-based analytics through sequence modeling tasks such as win-probability estimation, stroke forecasting, and movement prediction.
The dataset leverages broadcast video with expert labeling, forming a benchmark for research in action localization, report generation, and transfer learning.

ShuttleSet is a human-annotated stroke-level badminton singles dataset designed for tactical analysis and structured turn-based modeling. It was introduced as the largest publicly available badminton singles dataset with annotated stroke-level records, comprising 44 broadcast matches from 2018–2021, 104 sets, 3,685 rallies, and 36,492 strokes involving 27 top-ranking men’s singles and women’s singles players, and it couples per-stroke temporal, spatial, and tactical labels with match-, set-, and rally-level context (Wang et al., 2023). Subsequent work has treated ShuttleSet not only as a benchmark for badminton analytics, but also as a source domain for stroke-type recognition, doubles shot-occurrence transfer, temporal action localization, and sports table-to-text generation (Chang, 28 Feb 2025, Baek et al., 19 Aug 2025, Wang et al., 22 May 2026, Chiang et al., 29 Apr 2026).

1. Origin, scope, and corpus

ShuttleSet was created to address a specific deficit in turn-based sports analytics: the lack of public, expert-annotated, stroke-by-stroke datasets that preserve the full tactical structure of matches. The dataset focuses on badminton singles, where alternating strokes create strong short- and long-range dependencies across rallies and sets. Its source material is public broadcast video from bwf.tv rather than dedicated instrumented capture, which makes the dataset representative of realistic analysis conditions while retaining manual label quality (Wang et al., 2023).

The corpus spans quarter-finals, semi-finals, and finals of top-tier tournaments, including Super 1000 events and the BWF World Tour Finals, with a few World Tour Finals group-stage matches. The original release standardizes broadcasts to 30 frames per second and 1280×720 resolution. Match-level metadata include tournament name and level, round, year, duration, player names, and winner/loser; set-level metadata include scores; rally-level metadata include start, end, and winner; stroke-level metadata constitute the dataset’s core analytical unit (Wang et al., 2023).

A central feature of ShuttleSet is that it is structured around rallies rather than isolated clips. This makes it suitable for sequence modeling tasks such as win-probability estimation, next-stroke forecasting, and movement forecasting. The dataset was explicitly positioned as a resource for “turn-based analytics,” and its design reflects that orientation in both schema and benchmark construction (Wang et al., 2023).

2. Annotation schema, court representation, and shot taxonomy

Per stroke, ShuttleSet records four categories of annotations, expanded from the BLSR design. The rally category includes the current set score for both players before the stroke, the rally winner on the last stroke, and the lose reason, chosen from out, touched the net, not pass over the net, opponent’s ball landed, and misjudged. The temporal category includes hitting time in seconds and hitting frame. The spatial category includes player A and player B locations at the hitting moment, shuttle hitting location at contact, a categorical grid location, and a binary indicator of whether contact occurred above or below the net. The hitting category includes stroke status, backhand, around-the-head, and shot type (Wang et al., 2023).

The dataset stores both continuous court coordinates and categorical grid locations. Real-world court coordinates are obtained by homography, with $p' = H p$ , where $p$ is a 2D point in camera coordinates and $H \in \mathbb{R}^{3\times3}$ is estimated from court corners. Singles-court coordinates are defined with corners at $(0,0)$ , $(0,6.1)$ , $(13.4,0)$ , and $(13.4,6.1)$ . A domain-expert grid layout divides each half court into nine grids, plus seven outside grids, giving location features that are symmetric across court halves. ShuttleSet stores hitting and landing locations per stroke rather than continuous shuttle trajectories between frames (Wang et al., 2023).

The original shot taxonomy contains 18 stroke classes.

Classes 1–6	Classes 7–12	Classes 13–18
net shot, return net, smash, wrist smash, lob, defensive return lob	clear, drive, driven flight, back-court drive, drop, passive drop	push, rush, defensive return drive, cross-court net shot, short service, long service

These classes are accompanied by operational flags such as backhand and around-the-head, which encode distinctions that are difficult to infer reliably from raw video alone. The schema therefore mixes categorical tactics, court geometry, and event timing in a single per-stroke record (Wang et al., 2023).

3. Labeling workflow, quality control, and access

ShuttleSet was annotated with the S² (Shot-by-Shot) labeling tool, described as a microscopic labeling system optimized for badminton. The workflow has four stages: match preparation, rally segmentation, video preprocessing, and shot-by-shot labeling. During match preparation, videos are normalized to 30 fps and 720p, match metadata are entered, and court corners are marked for homography. Rally segmentation marks rally start/end and scoring with shuttle trajectory overlays as annotation aids. Video preprocessing splits full matches into rally clips. Shot-by-shot labeling then records the hitter location, opponent location, timing, contact context, return or landing time, landing location, and shot type in one pass through the rally (Wang et al., 2023).

Six trained annotators with badminton expertise performed the labeling. The paper reports random rally checks for quality assurance and re-labeling of three matches for consistency evaluation. Reported temporal consistency mean absolute error, in frames, was 0.24 for serve, 0.02 for return, and 0 for dead bird. Spatial consistency mean absolute error was 0.18 for shuttle position and 0.86 for player position. Shot-type consistency, expressed as accuracy difference upon re-annotation, was 0.00 for serve, 0.05 for return, and 0.11 for dead bird (Wang et al., 2023).

The dataset is distributed through the CoachAI project infrastructure. A Python implementation for loading and using ShuttleSet is provided through the CoachAI-Projects repository, and a Flask-based visualization platform supports match exploration through position analysis, shot-type analysis, and movement analysis. Over two years, this platform was used by national badminton teams during multiple international high-ranking matches, indicating that ShuttleSet was intended not only as a benchmark corpus but also as an operational analytics substrate (Wang et al., 2023).

4. Native benchmarks and baseline tasks

The original ShuttleSet paper defined three benchmark tasks. The first, stroke influence, estimates the rally winner from a stroke sequence and defines per-stroke influence as $I_t = P(\mathrm{win}\mid s_{1:t}) - P(\mathrm{win}\mid s_{1:t-1})$ . It uses AUC, accuracy, and Brier score as evaluation metrics, with the latest 10 matches reserved for test. The best reported model was ShuttleScorer, which achieved AUC 0.8371, ACC 0.7869, and BR 0.1574 (Wang et al., 2023).

The second task, stroke forecasting, predicts future shot types and shuttle landing locations from a prefix of observed strokes. It uses categorical cross-entropy for shot type and MAE/MSE for location, with prefixes of length $\tau \in \{8,4,2\}$ . The evaluation protocol generates 10 candidate sequences and scores the closest to ground truth. The best reported model was ShuttleNet; for $\tau=8$ , it achieved CE 2.4125, MSE 1.8121, and MAE 1.3582 (Wang et al., 2023).

The third task, movement forecasting, jointly predicts future strokes and both players’ future 2D positions. It uses the same train/test split as stroke forecasting and averages positional errors across both players and time. DyMF reported the best CE for $p$ 0 with 2.3146, while ShuttleNet reported the lowest MSE at 1.0832 and MAE at 1.5852 under the same horizon, indicating that explicit interaction modeling and position-aware fusion both matter, but not identically across metrics (Wang et al., 2023).

Task	Best reported model	Representative result
Stroke influence	ShuttleScorer	AUC 0.8371, ACC 0.7869, BR 0.1574
Stroke forecasting	ShuttleNet	$p$ 1: CE 2.4125, MSE 1.8121, MAE 1.3582
Movement forecasting	DyMF / ShuttleNet	$p$ 2: DyMF CE 2.3146; ShuttleNet MSE 1.0832, MAE 1.5852

These tasks collectively position ShuttleSet as a benchmark for structured sequential inference rather than only frame-level recognition. They also expose difficulty arising from the full 18-class taxonomy, class imbalance, and the dependence of tactical interpretation on both spatial and temporal context (Wang et al., 2023).

5. Repurposing in later research

Later studies have extended ShuttleSet well beyond its original benchmark suite. In skeleton-based stroke-type recognition, the "Badminton Stroke-type Transformer" paper uses ShuttleSet as the largest publicly available badminton video dataset for fine-grained stroke classification in singles. It leverages stroke labels, rally segmentation, hit frames, and court-line information; splits stroke classes by top-player versus bottom-player camera-view side; and reports that, after merging rare “driven flight” into “None,” the task is framed as 35 categories including “None.” On a 30/5/5 train/validation/test match split after removing problematic data, the best BST-3 variant with the proposed variable-width segmentation achieved Acc 0.7710, Macro-F1 0.7042, and Acc-2 0.9340, outperforming TemPose variants and generic skeleton action-recognition baselines (Chang, 28 Feb 2025).

A second line of work uses ShuttleSet as a source domain rather than an end benchmark. "Bridging the Gap: Doubles Badminton Analysis with Singles-Trained Models" trains a pose-based shot-occurrence representation on ShuttleSet singles videos using frame-level shot annotations only, converts them into “shot” versus “notshot” segments per player, and transfers the resulting models directly to doubles without fine-tuning. The paper reports singles accuracy 0.8617 for its pose-based Transformer at threshold 0.50, and doubles combined accuracy 0.6686 at threshold 0.86, compared with 0.5941 for a vision-only YOLOv11x baseline. The reported transfer degradation is ~19.3% for the pose-based model versus ~39.9% for the vision-only baseline, which the authors interpret as evidence that skeletal motion patterns learned from singles generalize better to doubles than raw visual cues (Baek et al., 19 Aug 2025).

A third reuse reframes ShuttleSet as a temporal action localization benchmark. "Decoupling Spatio-Temporal Adapter for Fine-Grained Badminton Action Localization" notes that ShuttleSet lacks temporal interval annotations and therefore constructs synthetic 19-frame segments centered on each annotated stroke frame, using $p$ 3. After resampling all videos to 25 fps, resizing frames to 160×160, and feeding VideoMAE-B features to an ActionFormer head with the proposed DSTA adapter, the method reports mAP 75.17 at tIoU 0.3, 75.10 at 0.4, 75.04 at 0.5, 74.84 at 0.6, and 73.18 at 0.7, for an average mAP of 74.67 (Wang et al., 22 May 2026).

ShuttleSet has also been transformed into a report-generation benchmark. "Tree-of-Text" introduces ShuttleSet+, a badminton match-reporting dataset derived from ShuttleSet22. ShuttleSet22 is described there as covering 58 matches played between 2018 and 2022, 140 sets, 3,992 rallies, and 33,612 strokes. ShuttleSet+ pairs multi-table match data with human-written English reports from Badminton World Federation and Olympics sources. On ShuttleSet+, Tree-of-Text reports RG # 16.78, RG P% 95.79, CS F% 78.03, CO DLD% 69.30, time 29.04, and cost 5.71, outperforming prompt-only baselines and using roughly 40% of the time and cost of Chain-of-Table on this benchmark (Chiang et al., 29 Apr 2026).

6. Limitations, label remapping, and research significance

Several limitations recur across ShuttleSet-based studies. The original dataset covers singles only, and the source paper explicitly notes that doubles introduces distinct tactical patterns and interactions not covered by the release. The original dataset also does not provide a fixed train/validation/test partition; later papers construct their own splits, sometimes by match and sometimes by rally or frame, which affects comparability across studies (Wang et al., 2023).

Downstream work also exposes task-dependent label remapping. The BST study removes problematic data, merges rare “driven flight” into “None,” and duplicates categories by top-player and bottom-player viewpoint. It further reports qualitative confusion between visually similar fine-grained actions such as top smash versus top wrist smash and bottom smash versus bottom wrist smash, as well as particularly poor recall for “top defensive return drive” (Chang, 28 Feb 2025). The TAL study, by contrast, retains the original 18-label taxonomy but must synthesize fixed temporal windows because ShuttleSet does not include action interval boundaries (Wang et al., 22 May 2026). This suggests that direct comparison across ShuttleSet papers requires careful attention to how labels, temporal units, and splits are redefined for each task.

The dataset’s broadcast-video origin is both a strength and a constraint. It supports realistic computer-vision benchmarking, but it also limits biomechanical precision: continuous shuttle trajectories are not stored, fine-grained contact semantics can be ambiguous in fast exchanges, and later work reports that 3D pose lifting can be unreliable for badminton-specific poses, making 2D pose sometimes preferable in practice (Chang, 28 Feb 2025). Class imbalance is another persistent issue: the original dataset reports rare categories such as defensive return lob, driven flight, and rush, and later work explicitly merges rare classes to stabilize training (Wang et al., 2023).

Despite these constraints, ShuttleSet occupies a distinctive position in badminton analytics. It is simultaneously a structured event dataset, a sequence-modeling benchmark, a source of training supervision for video models, and a basis for applied tactical interfaces. A plausible implication is that its long-term importance lies less in any single benchmark score than in the fact that it established a shared, stroke-level representation on top of public broadcast video, enabling otherwise disconnected research threads—forecasting, recognition, localization, transfer, and report generation—to operate on a common substrate (Wang et al., 2023).