Broadcast
Cuts, replays and close-ups break tracks, so IDs are re-linked after each cut and court keypoints re-solved per shot.
Basketball · Computer vision annotation
Basketball video annotation that keeps every ID.
Player tracking that survives screens and camera cuts, court keypoints, pixel masks, pose and box-score events. Labelled by people who know the game, checked three times.
NBA · NCAA · FIBA · HIGH SCHOOL · 3X3
TRACKING
ID LEDGERCUTS SURVIVED · 0
ID 1
ID 2
ID 3
ID 4
ID 5
ID 6
ID 7
ID 8
ID 9
ID 10
Illustrative simulation. Real labels follow your ID rules.
5
Label types, one supplier
3 layers
QA on every delivery
Native res
Masks at source resolution
Free
Pilot: one clip, up to about 10 minutes
What we label
Five kinds of label. Pick the layer.
Choose a label type to see what it looks like on a court and what lands in your delivery.
Every player boxed, every ID kept.
A box and a persistent ID for every player, held through screens, pile-ups and camera cuts, with team and jersey attached when they can be read.
- Bounding box per player per frame
- Persistent track ID across cuts and replays
- Team, role and jersey attributes
- Unknown state instead of guessed links
FORMATS · MOT, COCO video, JSON
TRACKING & RE-ID · EXAMPLEHOMEAWAY
Illustrative example drawn on a regulation NBA court.
The hard part
The hard part of basketball is identity.
Ten players in similar kit, screens on every possession, and a director cutting between cameras. A box can be perfect in every frame while the ID on it is wrong. So identity gets its own written rules before anyone labels.
Scope
How long an ID must hold: a shot, a clip, the full game.
Re-entry
Whether a player who leaves and returns keeps the same ID.
Evidence
Jersey, motion, position, appearance: ranked, written down.
Unknown
A real option, so nobody guesses after a pile-up.
- Read the full ID rules we publish
Sample output
Files your pipeline reads.
Boxes and IDs per frame, named court keypoints, class-index masks and timestamped events, in JSON, COCO, MOT or your own structure.
labels.json
{"frame": 1200,"shot_id": 3,"players": [{ "track_id": 1, "team": "home", "jersey": "23","bbox_xywh": [812, 402, 64, 168], "visible": true },{ "track_id": 6, "team": "away", "jersey": null,"jersey_note": "not readable","bbox_xywh": [874, 398, 61, 171], "visible": true }],"court_keypoints": [{ "name": "free_throw_center", "xy": [1182, 402], "visible": true }]}
Example structure with synthetic labels, not a real game. Your delivery follows your spec.
Example JSONHow it works
Video in. Labels out.
- 01
Send footage
Broadcast feeds, fixed gym cameras or practice video. A link or a bucket is enough.
- 02
Agree the spec
Classes, keypoints, ID rules and court template written down and versioned before anyone labels.
- 03
Label and check
An annotator labels, a basketball specialist reviews, a senior auditor signs off.
- 04
Receive data
Files in your format with a QA report, ready for your training pipeline.
Footage we handle
Any court. Any camera.
Each kind of basketball video fails in its own way. Here is what we do about it.
Fixed gym camera
Wide lenses and small players make numbers unreadable, so identity leans on team, position and movement.
3x3 and outdoor
Non-standard courts and changing light, so the court template and classes are fixed to your venue first.
Practice and drills
Repetitive movement suits pose and shot-form work, so keypoint rules are set for the motions you study.
Quality you can check
Errors caught before you train.
Most costly basketball label errors are swapped IDs after a screen and landmarks placed on the wrong line in a tight shot. Each layer looks for a different kind of mistake, and every delivery ships with a QA report.
How to check any vendor's labelsLAYER 1
Primary annotator
Labels every frame to the agreed spec: boxes, IDs, points, masks or events.
LAYER 2
Specialist reviewer
Checks basketball logic: the right player after a screen, the right landmark on a tight shot.
LAYER 3
Senior auditor
Audits a sample against the spec, including every camera cut, and signs off the QA report.
Pricing
Quoted per project. Tested for free.
PER PROJECT
Priced by task and volume
Boxes, keypoints, masks and events take different effort, so we quote after a short call and a look at your footage and spec.
How our pricing worksFREE PILOT
Your hardest possession
One clip, up to about 10 minutes, labelled with full QA and returned in 3 business days. No contract, no payment.
Questions
What CV teams ask us.
Can you keep player IDs across camera cuts?
Yes. We agree the ID rules first: how long an ID must hold, what evidence links two appearances, and when a player is marked unknown instead of guessed. Our player re-ID labeling guide sets out the rules we work to.
Read the re-ID labeling guideDo you deliver segmentation masks at native resolution?
Yes. Masks are single-channel class-index PNGs at the source resolution, using a class map you approve before work starts.
Can you use our own classes, keypoints or court template?
Yes. We work to your spec, ours, or a mix, and version it so a change mid-project never silently alters earlier frames.
Which courts do you handle?
NBA, NCAA, FIBA, high-school and 3x3 courts. Their dimensions differ, so the court template is fixed per project before keypoints are labelled.
How is basketball annotation priced?
Per project, by task, footage volume and label density. We quote after a short call and a look at your footage. New clients start with a free pilot.
Can you check labels another vendor produced?
Yes. We can review an existing dataset against your spec and report the errors we find before you train on it.
About our quality reviewFree pilot
Send us the clip your model gets wrong.
The screens, the scrums under the rim, the cut to a close-up. We label it and send the QA report.