Table 1.
Types of screen-plays in experiments and real games.
All types of screen-plays in this study are shown. Down screen, back screen, and pick dominated the offense in the controlled experiment but down screen, flare screen, cross screen, and pick dominated the offense in the real game. Experiment and real game indicate our two different datasets: a controlled experiment and real game, respectively.
Fig 1.
On- and off-ball screen-play movements.
(a-d) Screener’s position in on- and off-ball screen-plays. (a-b) Results in the controlled experiment. (c-d) Results in the real games.
Fig 2.
In this study, the classification was composed of three steps: data segmentation, extracting feature vectors and classifying plays. In segmentation, data was segmented into actions. After segmenting train data, the actions were labeled as screen-play or not. From segmented data of actions, feature vectors were extracted. By using these labels and feature vectors, the classifier was trained in a training session. After the training session, test session was executed with test data.
Fig 3.
Receiver operating characteristic (ROC) and recall-precision trade-off curve in screen detection.
ROC curves (top) and recall-precision trade-off (bottom) for on-ball (a and d), off-ball (b and e) and all screen-plays (c and f) are shown.
Fig 4.
Screen detection misclassification.
Examples of (a) false negative (miss) and (b) false positive (false-alarm) errors. Red, magenta, blue and orange lines represent the screener, user, defense of the user and the ball respectively. Triangles and circles represent their start positions. (a) Screener set a screen but the user moved far from screener so that this play was not classified as screen-play. (b) Two players merely moved closely and a screen-play was not set but this play was classified as screen-play incorrectly because of the distance between players.