BFMD
BFMD Dataset
A full-match dense badminton video dataset for shot captioning
BFMD
A full-match dense badminton video dataset for shot captioning
BFMD covers 19 full matches collected from four BWF World Tour Super-level events in 2025.
Anders Antonsen vs. Chou Tien Chen — F
Kodai Naraoka vs. Anders Antonsen — SF
Chou Tien-Chen vs. Shi Yu Qi — SF
Lee Chia Hao vs. Alex Lanier — SF
Each rally is annotated with four complementary modalities.
As a full-match dataset, BFMD provides dense temporal labels covering the entire broadcast. Each match is segmented into rallies, replays, Hawk-Eye challenges, and game intervals.
Per-frame bounding boxes for all players on court. Covers both singles and doubles matches across full-match sequences, enabling player detection and multi-object tracking research.
Full-body skeleton keypoints for each player. Captures precise limb and joint positions at every frame, supporting stroke classification, action recognition, and biomechanical analysis.
Frame-level shuttlecock positions from serve to landing. Captures full flight arcs, net passes, and landing coordinates, enabling trajectory prediction and rally structure analysis.
Natural language descriptions for each shot within a rally, covering shot type, direction, player position, and tactical intent. Designed to support video-language grounding and dense video captioning tasks.
“The player in white initiates play with an underhand stroke, striking the shuttle at a low contact height close to the service line to start the rally.”
“The player gently plays the shuttle from close to the net with a controlled, delicate touch, intending to land tightly in the front court and force the opponent forward.”
“The player in red executes an overhead clear from the rear court, sending the shuttle high and deep toward the opponent’s backcourt.”
“The player in the gray kit initiates the rally with an underhand motion from behind the service line, delivering the shuttle toward the opponent to start the point.”
“The player in red performs a delicate, controlled forehand net shot, causing the shuttle to travel tightly over the net and land close on the opponent’s side, prioritizing placement over power.”
Frame-level shot type annotations covering the full stroke vocabulary: serve, clear, drop, smash, drive, net shot, lift, push, and more. Enables fine-grained action recognition and tactical pattern mining across entire matches.
Player bounding boxes, skeleton poses, shuttle trajectory, and shot captions rendered together on a single rally clip.
Technical report describing the dataset collection pipeline, annotation protocol, and benchmark results.
Read on arXivBaseline models, evaluation scripts, and data loaders will be released on GitHub.
View on GitHubAccess is granted for academic research purposes only. Please fill out the request form with your name, institution, email, and a brief description of your research use case. Approved requests receive a download link by email.
Request AccessBy requesting access you agree that the dataset will be used solely for non-commercial academic research and will not be redistributed.