(Groupings of similarly frequent labels can be found here.) Instead, a model must be trained on between 9 and 11 classes with comparable label frequencies. Accordingly, training all 16 classes at once for video classification is not realistically doable.
NBA PLAY BY PLAY MP4
For instance, the NBA omitted mp4 clips of free throws for whatever reason. It should be noted that there is a huge discrepancy between the number of examples for certain classes.
Some of these classes include ASSISTED 2PT, 2PT MADE (unassisted), FIELD GOAL MISSED (marked as FGA), ASSISTED 3PT, 3PT MADE (unassisted), STEAL, OTHER TURNOVER, FOUL, and REBOUND. Although the full dataset download is quite large, users can abort the download early for a fully functional partial dataset and then easily restart the download from the stop point if more data is needed later.Įach clip can labelled into one of 16 possible classes. The label files will all contain an event classification/category, jersey number(s) of player(s) involved in the event, and the time remaining on the clock at the time of event. Each clip is about 10-15 seconds in length and has a corresponding XML label file. In its entirety, this dataset is over 1.75 TB with over 500,000 unique mp4 clips from approximately 2,500 NBA games over the past two seasons.
0 kommentar(er)
