Our Take
SAOT worked because MIT forced vendors to prove their data collection and algorithms could withstand real-world conditions—not because the technology itself was novel, but because the validation process was rigorous.
Why it matters
Sports organizations now rely on unproven vendor systems to make high-stakes decisions. This story shows what due diligence looks like when lives (or tournament outcomes) hang on computer vision accuracy.
Do this week
If you're deploying third-party tracking or detection systems in high-consequence settings, audit the raw skeletal or coordinate data before trusting algorithmic output—anatomically impossible positions are a baseline sanity check.
How MIT validated offside calls at the World Cup
The 2022 FIFA World Cup in Qatar was the first tournament to use semi-automated offside technology (SAOT) at scale. The system assisted in more than 150 offside calls across 64 games, overturning eight goals, adding two goals to the scoreboard that had been incorrectly disallowed, and changing the outcome of seven matches (per the article). The most visible application came in the Argentina v. France final, where SAOT confirmed that Argentine forward Lautaro Martinez was onside despite appearing to be in an illegal position.
The MIT Sports Lab, founded in 2015 by mechanical engineering professor Anette Hosoi and entrepreneur Christina Chase, played a critical role in validating this system. SAOT relies on skeletal data—3D representations of player positions—captured by 12 high-speed stadium cameras (per the article). That generates approximately 108,900 data points per second from 25 people on the field, each with 29 joints tracked across three dimensions. Add ball-tracking data from an embedded chip, and a single match produces more than a dozen gigabytes of raw data.
When FIFA acquired skeletal data from third-party vendors around 2021, the organization lacked in-house expertise to validate it. The Sports Lab was handed the problem. Immediately, the researchers found catastrophic errors: skeletons floating above ground, limbs stretching to impossible lengths, balls behaving in ways that violated physics. Henry Wang, a FIFA research consultant at the lab, recalls the raw feeds showed "anatomically impossible positions."
From 2021 through 2022, FIFA rented out stadiums for multi-day test sessions. Lab researchers analyzed offside drills performed by amateur players and FIFA staff while vendors collected live data. The Sports Lab's job was to measure latency (how fast data could be delivered), validate the skeletal output against ground truth, and test whether combining skeletal and ball-tracking data could produce reliable offside calls. The team developed a Google Cloud tool to capture data in real time and establish protocols to synchronize the two data streams.
Validation is the unglamorous work that makes high-stakes automation safe
SAOT is not a breakthrough in computer vision or skeletal tracking. The underlying techniques—multi-camera pose estimation and ball tracking—were already mature. What made SAOT deployable was the willingness to run dozens of test drills, measure errors in anatomically impossible ways, and push vendors to fix blind spots they did not know they had.
In soccer, a single offside call can determine a championship. The stakes are lower elsewhere but the principle is identical: organizations buying third-party tracking, detection, or classification systems often lack the technical expertise to validate vendor claims. SAOT worked not because MIT invented anything new, but because someone forced the vendors to prove their systems worked before they went live on the world's biggest stage.
Ferran Vidal-Codina, a former MIT research scientist who led the validation effort, notes that "decisions have been made quicker and better." The system also shifts accountability. Referees still make the final call, but now they see algorithmic evidence. During World Cup play, animated visuals were shown to stadium audiences and up to 5 billion viewers on broadcast platforms.
Before deploying vendor systems, demand raw-data audits
If you are integrating a third-party tracking, pose estimation, or detection system into a high-consequence workflow—sports, medical imaging, autonomous systems, compliance monitoring—do what MIT did: obtain raw output (skeletal coordinates, detection boxes, confidence scores) and run sanity checks before trusting the algorithm's final judgment.
Anatomically impossible positions (joints stretching beyond physiological limits), temporal inconsistencies (objects jumping between frames), and edge-case failures in controlled test scenarios are early-warning signs that the vendor's preprocessing or model has not been adequately tested. Ask the vendor to re-train or recalibrate on data that matches your deployment environment. Do not assume that accuracy metrics published in a paper or datasheet will hold in live conditions.