Considering the enormous creation rate of usergenerated videos on websites like YouTube, there is an immediate need for automatic categorization, recognition and analysis of videos. To develop algorithms for analyzing user-generated videos, unconstrained and representative datasets are of great significance. For this purpose, we collected a dataset of Sports Videos in the Wild (SVW), consisting of videos captured by users of the leading sports training mobile app (Coach's Eye) while practicing a sport or watching a game. The dataset contains 4100 videos selected by reviewing ~85,000 videos and consists of 30 sports categories and 44 actions. Videos of sports practice, which frequently happens outside the typical sports field, have huge intra-class variations due to background clutter, unrepresentative environment, existence of different training equipment and most importantly, imperfect actions. On the other hand, using smartphones for video capturing by ordinary people, in comparison to videos captured by professional crew for broadcasting, leads to challenges due to camera vibration and motion, occlusion, view point variation, and poor illumination. Given various manual labels, this dataset can be used for a wide range of computer vision applications, such as action recognition, action detection, genre categorization, and spatio-temporal alignment. On the sport genre categorization problem, we design the evaluation protocol and evaluate three different methods to provide baselines for future works.
展开▼
