Semi-supervised Learning for Identifying Players from Broadcast Sports Videos with Play-byPlay Information

Tracking and identifying players in sports videos filmed with a single moving pan-tilt-zoom camera has many applications, but it is also a challenging problem due to fast camera motions, unpredictable player movements, and unreliable visual features. Recently, [26] introduced a system to tackle this problem based on conditional random fields. However, their system requires a large number of labeled images for training. In this paper, we take advantage of weakly labeled data in the form of publicly available playby-play information. This, together with semi-supervised learning, allows us to train an identification system with very little supervision. Experiments show that by using only 1500 labels with the play-by-play information in a dataset of 75000 images, we can train a system that has a comparable accuracy as a fully supervised model trained by using 75000 labels.