Sports video annotation & tactics
Tag events and measure performance from match video.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Hudl · Kinovea · Dartfish
- Class GUI-need label
G3· GUI-only
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R13.05.01.01-T01 Measure knee angle at foot strike in a running video with Kinovea
Original instruction
Open input/treadmill_side_240fps.mp4 in Kinovea. Step to the first right-foot strike after 00:02.000, place an angle tool on hip-knee-ankle markers (stickers visible on the runner) and record the knee angle. Save the Kinovea annotation file (.kva) and write the angle and frame number in answer.json.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometryav· Audio/video over timevis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Frame number must equal the labelled foot-strike frame ±2; angle within ±3° of the ground truth from marker coordinates; .kva contains an angle drawing on that frame whose 3 points lie within 6 px of the markers. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the .kva, so a file not written by Kinovea is detected.
Proposed partial credit
30% correct frame; 40% angle within 3°; 30% annotation points on markers
R13.05.01.01-T02 Log every service ace in a tennis match video
Original instruction
input/tennis_set1.mp4 (40 min, one set) is loaded in LongoMatch with an empty project. Create an event type 'Ace' and tag every service ace (serve not touched by the receiver) with the event start at ball contact, then export the timeline as CSV.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timeclick· Point & clicknav· Pan/zoom/scroll- Work product
WP10· Annotations- Conditions
long· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Exported CSV events matched to ground truth aces within ±1.0 s: precision and recall. Service winners touched by the receiver are hard negatives. Project DB edits outside LongoMatch are detected by its revision log.
Proposed partial credit
60% recall; 30% precision; 10% event type and export correct
R13.05.01.01-T03 Count completed passes per team in a football clip
Original instruction
input/football_10min.mp4 is a broadcast clip (10 minutes). Using LongoMatch, tag every completed pass for each team (event types 'Pass Home' and 'Pass Away', tagged at the pass moment), then report totals in answer.json.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timeclick· Point & clickvis· Visual judgment- Work product
WP10· Annotations- Conditions
long· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Event lists matched to ground truth within ±1.5 s per team (F1); answer.json totals within ±5% of truth for full credit. The agent has no shell, Python or ffmpeg, and the process writing the project DB is logged (inotify/FSEvents/Sysmon), so events not tagged in LongoMatch are detected.
Proposed partial credit
60% event F1 (both teams); 40% totals within tolerance
R13.05.01.01-T04 Build a tagging dashboard with hotkeys and code the first half
Original instruction
In LongoMatch, create a dashboard 'Hockey Basic' with 8 event buttons (Shot, Goal, Faceoff Won, Faceoff Lost, Turnover, Penalty, Save, Hit), each with a distinct hotkey and team-tag support. Then code the first 10 minutes of input/hockey_p1.mp4 for Shot and Goal events only.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clickkey· Shortcutsav· Audio/video over time- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Project/dashboard JSON: 8 buttons with exact names, unique hotkeys. Shot and Goal events in 0-10:00 matched to ground truth ±1.5 s with team (F1). The agent has no shell, Python or ffmpeg, and the process writing the project files is logged (inotify/FSEvents/Sysmon), so a dashboard or events not made in LongoMatch are detected.
Proposed partial credit
40% dashboard spec; 60% scaled by Shot/Goal F1
R13.05.01.01-T05 Calibrate distance and measure a 10-20 m sprint split in Kinovea
Original instruction
input/sprint_side.mp4 shows a sprinter passing cones at 10 m and 20 m (a 1 m calibration bar is on the ground). In Kinovea, calibrate with a line on the bar, track the hip marker with the trajectory tool, and measure the 10-20 m split time and mean speed. Save the .kva and answer.json.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometrydrag· Drag & dropav· Audio/video over time- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Split time within ±0.02 s and speed within ±2% of ground truth (from synchronized timing gates). .kva contains a calibration line of 1 m and a trajectory spanning the segment. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the .kva, so a file not written by Kinovea is detected.
Proposed partial credit
25% calibration; 25% trajectory tracked; 25% split time; 25% speed
R13.05.01.01-T06 Synchronise front and side camera videos on a key event
Original instruction
In Kinovea's dual-screen mode, open input/jump_front.mp4 and input/jump_side.mp4 (different start times). Synchronise them on the take-off frame of the countermovement jump, then export a side-by-side video jump_sync.mp4 covering 1 s before to 1 s after take-off.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timenav· Pan/zoom/scrollvis· Visual judgment- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
jump_sync.mp4: duration 2.0 ±0.1 s; frame-difference analysis of each half shows take-off at the centre frame in both (±1 frame at 120 fps), giving sync error ≤ 2 frames. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the export, so a video not exported by Kinovea is detected.
Proposed partial credit
50% sync error within 2 frames; 30% clip window; 20% side-by-side export
R13.05.01.01-T07 Draw a golf swing plane line and measure shaft angle at the top in Dartfish
Original instruction
Open input/golf_dtl.mp4 (down-the-line view) in Dartfish. At address, draw the swing-plane line along the shaft; at the top of the backswing, measure the shaft angle relative to the plane with the angle tool. Export the annotated key frame as PNG and record the angle in answer.json.
- Platforms
win· Windowsmac· macOS- Applications
- Dartfish
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometrypath· Continuous pathav· Audio/video over time- Work product
WP10· Annotations- Conditions
licensed· Licensed pro app- Proposed difficulty
- 3 / 5
Proposed verifier
PNG: line drawn within 3° of the labelled shaft line at address (line detection on the overlay colour); answer angle within ±4° of truth; key frame at top of backswing ±3 frames (pHash). The agent has no shell, Python or ffmpeg, and the process writing the PNG is logged (FSEvents/Sysmon), so an image not exported by Dartfish is detected.
Proposed partial credit
30% plane line accurate; 30% correct key frame; 40% angle within tolerance
R13.05.01.01-T08 Code a basketball possession timeline in Hudl Sportscode
Original instruction
In Hudl Sportscode (macOS), open input/bball_q1.mov (one quarter) with the provided code window 'Possessions'. Code every possession for both teams with start at gain of possession and end at loss/score, and add labels for outcome (score, turnover, miss-rebound). Export the timeline XML.
- Platforms
mac· macOS- Applications
- Hudl Sportscode
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timeclick· Point & clickkey· Shortcuts- Work product
WP10· Annotations- Conditions
licensed· Licensed pro applong· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Timeline XML instances matched to ground truth by temporal IoU ≥ 0.6 per possession; label accuracy for matched instances; team correct. The agent has no shell or text editor, and FSEvents logs the process writing the timeline XML, so a file not written by Sportscode is detected.
Proposed partial credit
50% possession F1; 30% outcome labels; 20% team attribution
R13.05.01.01-T09 Track a barbell path and export trajectory data in Kinovea
Original instruction
input/snatch_side.mp4 shows a snatch with a coloured plate centre. In Kinovea, calibrate using the 450 mm plate diameter, track the plate centre through the lift and export trajectory data to CSV. Report the maximum horizontal deviation from the start position in cm.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
geom· Precise geometrydrag· Drag & dropav· Audio/video over time- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Exported CSV trajectory vs ground-truth track: mean point error ≤ 1.5 cm after calibration; reported max deviation within ±1 cm of truth. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the CSV, so a trajectory not exported by Kinovea is detected.
Proposed partial credit
25% calibration; 45% scaled by trajectory error; 30% max deviation
R13.05.01.01-T10 Tag corner kicks and their outcomes over a full half
Original instruction
input/football_half.mp4 (45 min) in LongoMatch. Tag every corner kick with team and an outcome sub-tag (shot within 10 s, cleared, possession retained), then export the event list as CSV.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timeclick· Point & clicknav· Pan/zoom/scroll- Work product
WP10· Annotations- Conditions
long· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Events matched to ground truth ±2 s: F1; outcome sub-tag accuracy on matches. The agent has no shell, Python or ffmpeg, and the process writing the project DB and CSV is logged (inotify/FSEvents/Sysmon), so events not tagged in LongoMatch are detected.
Proposed partial credit
60% corner F1; 30% outcome accuracy; 10% team accuracy
R13.05.01.01-T11 Draw an offside line telestration on a freeze frame
Original instruction
input/goal_clip.mp4 contains a disputed goal. In Kinovea, freeze the frame when the pass is played (ball leaves the passer's foot), draw a line parallel to the pitch markings through the second-last defender, mark the attacker, and export the annotated frame as offside.png.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
path· Continuous pathgeom· Precise geometryav· Audio/video over time- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Exported frame equals the labelled pass frame ±1 (pHash); line angle within 2° of the vanishing direction of the pitch lines and passing within 8 px of the defender's labelled foot. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing offside.png, so an image not exported by Kinovea is detected.
Proposed partial credit
40% correct frame; 40% line geometry; 20% attacker marked
R13.05.01.01-T12 Count stroke rate per 25 m from a swimming video
Original instruction
input/swim_50free.mp4 shows a 50 m freestyle from a side camera with lane ropes marked every 5 m. Using Kinovea's stopwatch and frame stepping, compute stroke rate (cycles/min) for each 25 m length and write strokes.csv.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timenav· Pan/zoom/scrollvis· Visual judgment- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
strokes.csv: two lengths with stroke rate within ±2 cycles/min of ground truth (from hand-entry labels). Python, ffmpeg and PowerShell are blocked by AppLocker, so the video cannot be analysed by script and strokes must be counted with Kinovea's frame stepping and stopwatch.
Proposed partial credit
50% first length within tolerance; 50% second length within tolerance
R13.05.01.01-T13 Grade 20 volleyball serve receptions on a 0-3 scale
Original instruction
In LongoMatch, open input/volley_set.mp4 with event type 'Reception' and labels 0-3 already defined. Tag each of the 20 serve receptions by the home team at contact time and label quality per input/reception_scale.pdf (0 = error, 3 = perfect pass to setter zone).
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
av· Audio/video over timeclick· Point & clickvis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Events matched to ground truth ±1 s; label agreement with coach key (exact = full, ±1 = half); quadratic-weighted kappa reported. The agent has no shell, Python or ffmpeg, and the process writing the project DB is logged (inotify/FSEvents/Sysmon), so events not tagged in LongoMatch are detected.
Proposed partial credit
40% event recall; 60% label agreement
R13.05.01.01-T14 Measure ground contact times for five sprint steps
Original instruction
input/sprint_240fps.mp4 shows 5 sprint steps from the side. In Kinovea, mark touch-down and toe-off frames for each step with key images, and report each ground contact time in ms in contacts.csv.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timenav· Pan/zoom/scrollvis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
contacts.csv: 5 values each within ±8.3 ms (2 frames) of ground truth from force plate sync; .kva key images exist at the chosen frames. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the .kva, so key images not made in Kinovea are detected.
Proposed partial credit
80% scaled by steps within tolerance; 20% key images saved
R13.05.01.01-T15 Build a goals highlight playlist and export clips
Original instruction
LongoMatch project input/match.lgm already has tagged events. Create a playlist 'Goals' containing every Goal event in chronological order, each clip from 8 s before to 4 s after the tag, and render the playlist to goals.mp4.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clickdrag· Drag & dropav· Audio/video over time- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Playlist in the project DB: all Goal events in order with lead/lag 8/4 s; goals.mp4 duration = 12 s × n ±0.5 s and each segment's start frame matches the source (pHash). The agent has no shell, Python or ffmpeg, and the process writing the DB and goals.mp4 is logged (inotify/FSEvents/Sysmon), so outputs not made by LongoMatch are detected.
Proposed partial credit
40% playlist content and order; 30% clip windows; 30% rendered video
R13.05.01.01-T16 Measure tennis ball toss height with calibration
Original instruction
input/serve_side.mp4 shows 6 serves with a 1.0 m reference pole beside the server. In Kinovea, calibrate, then measure the apex height of the ball toss above the ground for each serve. Write toss.csv with serve number and height in metres.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometryav· Audio/video over timevis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
toss.csv: 6 heights each within ±0.08 m of ground truth (from 3D-calibrated reference). Calibration line must exist in .kva. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the .kva, so a calibration not made in Kinovea is detected.
Proposed partial credit
20% calibration; 80% scaled by serves within tolerance
R13.05.01.01-T17 Annotate a basketball shot chart from game video
Original instruction
input/bball_half.mp4 (one half) and the court diagram template in Nacsport's drawing module. Tag every field-goal attempt by the home team with made/missed and place its location on the court diagram. Export events with coordinates.
- Platforms
win· Windows- Applications
- Nacsport
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
av· Audio/video over timegeom· Precise geometryclick· Point & click- Work product
WP10· Annotations- Conditions
licensed· Licensed pro applong· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Exported events matched to ground truth shots ±2 s; made/missed accuracy; court location error (after mapping diagram to court metres) ≤ 1 m for full credit. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing the export, so events not exported by Nacsport are detected.
Proposed partial credit
40% shot recall; 20% made/missed; 40% scaled by location error
R13.05.01.01-T18 Score difficulty elements of an aerobics routine from video
Original instruction
input/aerobics_trio.mp4 is a competition routine. Using input/code_of_points_excerpt.pdf, identify each difficulty element performed, its family and value, mark each with a timed event in LongoMatch, and write the total difficulty score to answer.json.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- LongoMatch
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
av· Audio/video over timevis· Visual judgmentclick· Point & click- Work product
WP10· Annotations- Conditions
long· Long-horizontutorial· Follow docs/tutorial- Proposed difficulty
- 5 / 5
Proposed verifier
Element list matched to judges' ground truth (time ±2 s, element code exact); total difficulty within ±0.1 of the official value. The agent has no shell, Python or ffmpeg, and the process writing the project DB is logged (inotify/FSEvents/Sysmon), so events not marked in LongoMatch are detected.
Proposed partial credit
50% element F1; 20% element values; 30% total score
R13.05.01.01-T19 Tag play types and filter third-down plays in Hudl
Original instruction
In the Hudl web app (test team account), the game film 'Week 3 vs Lakers' has 64 plays with Down and Distance filled. Add the play-type tag (Run/Pass) to each play from input/play_types.csv, then create a playlist of all 3rd-down pass plays.
- Platforms
web· Any OS (web)- Applications
- Hudl
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsnav· Pan/zoom/scroll- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Hudl test account state (export of the breakdown data): PLAY TYPE column equals the CSV; playlist contains exactly the 3rd-down pass plays. The test account has no API key and Hudl offers it no public API, so tags and playlists can only be made in the browser session.
Proposed partial credit
60% play types correct; 40% playlist set exact
R13.05.01.01-T20 Mark three events with key images and export them
Original instruction
In Kinovea, open input/long_jump.mp4 and add key images (with titles) at the last step before take-off, take-off and landing. Export the key images as JPEGs named with their titles.
- Platforms
win· Windows- Applications
- Kinovea
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timenav· Pan/zoom/scrollclick· Point & click- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Exported JPEGs: 3 files with the required titles; each perceptually matches a frame within ±3 frames of the labelled event. Python, ffmpeg and PowerShell are blocked by AppLocker, and Sysmon logs the process writing each JPEG, so images not exported by Kinovea are detected.
Proposed partial credit
33% pre-take-off step image; 33% take-off image; 34% landing image
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.