SoccER: Computer graphics meets sports analytics for soccer event recognition

Automatic event detection from images or wearable sensors is a fundamental step towards the development of advanced sport analytics and broadcasting software. However, the collection and annotation of large scale sport datasets is hindered by technical obstacles, cost of data acquisition and annotation, and commercial interests. In this paper, we present the Soccer Event Recognition (SoccER) data generator, which builds upon an existing, high quality open source game engine to enable synthetic data generation. The software generates detailed spatio-temporal data from simulated soccer games, along with fine-grained, automatically generated event ground truth. The SoccER software suite includes also a complete event detection system entirely developed and tested on a synthetic dataset including 500 minutes of game, and more than 1 million events. We close the paper by discussing avenues for future research in sports event recognition enabled by the use of synthetic data.