You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

SoccerTrack v2

Ten university-level soccer matches (934 minutes) recorded by fixed panoramic camera systems whose field of view spans the entire pitch, released together with frame-level game state annotations and player-linked ball action events on the same footage.

Contents

Folder Contents
videos/ 20 panoramic half-match videos (two 45-minute periods per match)
gsr/ Game state annotations: one JSON per half (pitch coordinates, jersey numbers and persistent player identifiers, roles, team sides)
bas/ Ball action events: twelve classes per match, with acting team and actor identifiers shared with the game state annotations; position is milliseconds from the first decoded frame of the half's video, and frame is the video (and game state) frame the event falls on
ball/ Ball track per half, one row per game state frame (auxiliary annotation). It is a reconstruction rather than a per-frame measurement: it is anchored at the annotated events and interpolated in straight constant-speed lines between anchors, so it carries event timing. Positions are clamped to the pitch rectangle in matches 132831 and 132877
mot/ Curated bounding-box subset used for the SoccerTrack Challenge at MMSports 2025: per-match ground-truth files and the paired 4-minute clips
raw/ Per-match calibration files (65 hand-annotated pitch keypoints per match, distortion maps, camera intrinsics) and the provider's raw tracking and event exports, with personal names removed

Version

This is release v1.2 (tag v1.2); earlier releases remain available as tags v1.1 and v1.0. In v1.2 every annotation lines up with the released video frame for frame: game state frame k is video frame k in every half, ball action events carry their frame, and times are milliseconds after the first decoded frame of each half's video. The unfilmed third period that three matches played, and the few events that fall outside the filmed frames, are removed, leaving 21,428 ball action events. Ids are unique across all twenty halves, so halves can be loaded together. CHANGES_v1_2.md lists every change; the main fix over v1.1 is the synchronisation of match 132877 (M10), whose positions were about one second out of step with its video.

Match labels

Files are named by the provider's match identifiers. The paper labels the matches M1 to M10 in ascending identifier order:

Paper label Match id Split
M1 117092 train
M2 117093 validation
M3 118575 train
M4 118576 train
M5 118577 train
M6 118578 train
M7 128057 test
M8 128058 train
M9 132831 test
M10 132877 validation

Benchmark split (SoccerTrack Challenge 2025): test M7 + M9, validation M2 + M10, train the remaining six. The paper's ball action spotting results use five-fold cross-match validation in which fold 0 tests on M7 + M9.

Coordinate convention

Pitch coordinates are metric on a 105 m x 68 m pitch with the origin at the centre circle; x increases towards the right half of the pitch as seen from the main camera and y towards the touchline nearer the camera.

Code and documentation

Loaders, evaluation CLIs (GS-HOTA, BAS mAP, MOT HOTA) and baseline starter kits: https://github.com/AtomScott/SoccerTrack-v2

License

CC BY 4.0. The release is pseudonymised: players are identified by jersey numbers and integer identifiers, and no personal names are included. Please cite the SoccerTrack v2 paper (in submission; citation to be added) when publishing results.

Downloads last month
2,020