Semi-supervised Learning for Identifying Players from Broadcast Sports Videos with Play-by-Play Information
By Jo-Anne Ting, UBC Computer Science
Abstract:
Tracking and identifying players in sports videos filmed with a single
moving, zooming camera has many applications, but it is also a
challenging problem due to fast camera motions, unpredictable player
movements, and unreliable visual features. We previously introduced a
system to tackle this problem based on conditional random fields.
However, that system requires a large number of labeled images for
training. In our most recent work, we take advantage of weakly labeled
data in the form of publicly available play-by-play information. This,
together with semi-supervised learning, allows us to train an
identification system with very little supervision. We also propose a
more robust way to predict identities of players at test time by using
a simpler model based on tracklets. Our results show that we can get
better identification results by using far fewer labeled training
examples; semi-supervised learning with only 150 labels in a
75000-image training set outperforms a fully-supervised model learned
on a 30000-image training set