Comprehensive Dataset of Broadcast Soccer Videos

In the absence of a rich dataset, sports video analysis remains a challenging task. Shot analysis, event detection, and player tracking are important aspects of soccer game research, however currently available datasets focus mainly on single-type annotations, and have several limitations when applied to full video analysis; these include a lack of training samples, temporal annotations, and rich labels of different types. In this paper, we focus on broadcast soccer videos and present a comprehensive dataset for analysis. In its current version, the dataset contains 222 broadcast soccer videos, totaling 170 video hours. The dataset covers three annotation types: (1) A shot boundary with two shot transition types, and shot type annotations with five shot types; (2) Event annotations with 11 event labels, and a story annotation with 15 story labels at coarser granularity; and (3) bounding boxes of the players under analysis in a subset of 19908 video frames. We hope that the introduction of this dataset will enable the research community to develop soccer video analysis techniques.