Detecting zone defense & measuring its impact

The Zone Defense Detection Model will give fans a deeper understanding of team strategy, as well as offensive and defensive performance.

Determining when defenses are in zone or man has long been something coaches and fans could do with the eye test. However, labeling the defense of every possession of every game, across multiple seasons, is another matter. This season, we’re introducing a Zone Defense Detection Model trained using Amazon Sagemaker, which will allow us to label every halfcourt possession as zone or man, giving us new, exciting ways to understand game strategy.

In simple terms, zone defense is when each player on defense is assigned an area of the floor to defend, while in man-to-man each defender guards a specific player. The two can look similar for stretches of a possession; however, the model is trained to account for subtle differences in how players move, using the player tracking data to determine if a team is playing zone defense.

The classification of each possession into either zone or man defense is driven by the NBA’s optical tracking system, which uses 3D pose detection to track 29 points on every player’s body, 60 times per second. These detailed body-position coordinates feed advanced Machine Learning models, allowing every movement and spatial relationship on the floor to be captured.

Like a fan watching the game, the model follows all five defenders across an entire possession, tracking how they move in relation to the ball, the offense, and each other to identify patterns. In man-to-man, defenders more often follow their assignments wherever they go; in a zone, they hold their areas and hand cutters off to the next defender. The model learned those movement signatures from thousands of possessions, and now it applies that knowledge to every possession of every game, labeling each one as zone or man. That label is the foundation for everything else.

By classifying each possession, fans will have a deeper understanding of team strategy, as well as offensive and defensive performance. This includes how efficiently opponents shoot when a team uses zone versus man, how often and in what game situations teams turn to zone, and how team offenses fares against each coverage. Zone defense is rare, run on roughly 2.5% of all possessions league-wide, but a handful of teams lean on it heavily, and now fans can see the exact results when it’s run.

Identifying zone defense is part of the next generation of stats and insights that we’re bringing to fans in real-time with the power of AWS Machine Learning.