Authors:
Carlos Olea
1
;
Gus Omer
2
;
John Carter
2
and
Jules White
1
Affiliations:
1
Dept. of Computer Science, Vanderbilt University, 1025 16th Ave. S, Nashville, U.S.A.
;
2
NOAH Basketball, 26958 US Highway 72, Athens, U.S.A.
Keyword(s):
Sports Analytics, Action Recognition, Computer Vision, Deep Learning.
Abstract:
Recent technologies have been developed to track basketball shooting and provide detailed data on shooting accuracy for players. Further segmenting this shot accuracy data based on shot types allows a more detailed analysis of player performance. Currently this segmentation must be performed manually. In this paper, we apply a state-of-the-art action recognition model to the problem of automated shot type classification from videos. The paper presents experiments performed to optimize shot type recognition, a unique taxonomy for the labeling of shot types, and discusses key results on the task of categorizing three different shot types. Additionally, we outline key challenges we uncovered applying current deep learning techniques to the task of shot type classification. The NOAH system enables the capture of basketball shooting data by recording every shot taken on a court along with its shooter and critical statistics. We utilized videos of 50,000 practice shots from various players
captured through NOAH system to perform the task of classifying shot type. These three second video clips contain the shooting action along with movements immediately preceding and following the shot. On the problem of shot type classification, the Temporal Relational Network achieved an accuracy of 96.8% on 1500 novel shots.
(More)