Overview
Our client is a US sports analytics company. Its analysts study team strategy and support player recruitment in American football, tennis, soccer and ice hockey, and they needed game footage turned into detailed event data: who did what, where and when. The work covers box scores and event tracking data across these sports. Over the project we annotated 550 games and tracked more than 40 player event types, which gives the client's models and analysts the data they work from.
Solution
A small team of annotators, ML engineers, a data scientist and an MLOps engineer set up the annotation process together. Custom annotation tools and machine learning models do the first pass on player tracking and event recognition, and annotators confirm and correct it. For ice hockey we built a workflow designed to deliver match analytics within 10 minutes of the final whistle.
Technology Stack
Features
- Multi-sport annotation: one process covering American football, tennis, soccer and ice hockey.
- Detailed event tracking: more than 40 player event types, such as passes, shots and ball handling, plus player positions on each frame.
- High volume: 550 games processed and annotated.
- Fast turnaround: the hockey workflow is designed to deliver analytics within 10 minutes after a game.
- Coverage around match times: work is scheduled around the fixture calendar.
- Model-assisted labelling: models propose positions and events, and annotators confirm and correct them.
- Efficient tooling: custom tools and workflows reduce the manual effort per game.
- Game analytics support: tracking and event recognition feed the client's analytics products.
- Strategy and recruitment data: the annotated events support the client's work on team strategy and player recruitment.
- Quality control: written guidelines and review steps keep annotations consistent across sports.

Outcome
- Detailed, consistent event data across four sports
- Player performance and game dynamics available to analysts soon after each game
- Data the client uses in its decisions on team strategy and player recruitment
- A process that scales up and down with the sports calendar
- Post-game analytics fast enough for the client to act on them the same day
games annotated
player event types tracked
from the final whistle to hockey match analytics
How We Organize Annotation at This Scale
Annotation quality is decided before the first frame is labelled. Our large annotation programs are built on the same components:
- An event taxonomy with edge cases. Every event type has a written definition, a start and end rule, and examples of what does not count. Most disagreements between annotators trace back to a definition, not to carelessness.
- Model-assisted labelling. Detection and tracking models propose player positions and candidate events; annotators confirm and correct. People spend their time on judgement, not on drawing boxes.
- Two-pass review with a gold set. A share of every batch is labelled twice and compared, and a hidden set of expert-labelled clips measures each annotator's accuracy continuously.
- Agreement tracking. Inter-annotator agreement is reported per event type and per sport. A falling score triggers a guideline update or retraining before the errors reach the client's models.
- A pipeline built for the clock. Delivering analytics minutes after the final whistle means annotating during the game, in parallel segments, with automated consistency checks merging the results.
Where a client needs results during live fixtures, shifts are planned around the match calendar and an experienced annotator resolves edge cases as they come up.
Why Sports Events Are Hard to Label
Sports footage looks unambiguous until two people label the same clip. When the authors of the SoccerNet-v2 benchmark measured how well human annotators agree with each other on action spotting, the result was about 80% on the benchmark's own metric. Trained people disagree on roughly one event in five when the definitions leave room for interpretation.
The causes are the same across sports. Broadcast cameras cut away at the moment of the action. Players occlude each other and the ball. Many events have no sharp boundary: the exact frame where a pass begins or a tackle ends is a convention, not an observation. Each sport adds its own difficulties, from the speed of the puck in ice hockey to simultaneous movement of every player in American football.
This is why consistent guidelines and measured agreement matter more than raw team size. A model trained on inconsistent labels learns the inconsistency. See our data annotation services and computer vision development pages for how annotation fits into a full model pipeline.