A Video Dataset of Shuttlecock-to-Ground Contact Events for Machine Learning-Based Detection in Badminton

This dataset contains 396 annotated high-speed video clips of badminton shuttlecock landings, comprising 41,982 frames and 573 frame-level annotated shuttlecock-to-ground contact events. It was built for frame-accurate detection of the moment a shuttlecock touches the ground — the decision an electronic line-calling system in badminton has to make — and is, to the authors' knowledge, the first publicly available dataset designed specifically for this task. Each clip is a single shuttlecock flight recorded by a ground-level monochrome industrial camera (Basler ACE, 800 × 600 px) aimed along a court line and mounted rotated by 90°, as part of the FlyEye Challenge System. Clips are released as sequences of 136 × 136 pixel 8-bit grayscale PNG patches centred on the shuttlecock, together with one XML annotation file per clip giving the shuttlecock position in every frame, in source-frame coordinates, and the indices of the frames in which contact with the ground occurs. The material was recorded in 2018–2019 at two Polish venues: Kahuna Sports Club in Warsaw (344 clips, 150 fps) and the Spodek Arena in Katowice during the BWF World Senior Championships 2019 (52 clips, 190 fps). The Katowice clips retain visible lighting flicker; this is deliberate, as the released patches are cropped from raw frames rather than from the de-flickered stream used internally for shuttlecock detection. An official split into training, validation and test subsets (276 / 40 / 80 clips), defined at the level of whole clips, is part of the dataset so that results are directly comparable across studies. Contact frames are ambiguous at the one-frame scale — two annotators picked the same frame in 80.6% of clips and agreed within one frame in 99.5% — so evaluation should be reported both strictly and with a tolerance of one frame. The archive includes a full datasheet (README.md) documenting the annotation format, recording setup, statistics, evaluation protocol and known limitations, the licence text, machine-readable citation metadata, a per-clip metadata index, and SHA-256 checksums for every file.