SportLight: statistically principled crowdsourcing method for sports highlight selection

Sports highlight selection has traditionally required expert opinions and manual labor of video editors. To automate this laborious task, crowdsourcing viewers’ live comments has recently emerged as a promising tool, which can remove the burden of extracting semantic information by computer vision. However, popular crowdsourcing methods based on peak-finding are sensitive to noise and may produce deviant highlights from the expert choice. To increase the accuracy of automated selection of sports highlight, we introduce a statistically sound crowdsourcing method, SportLight. In this work, we take a statistical approach that combines multiple hypothesis testing and ℓ 1 -trend filtering (fused lasso), supported by a computationally inexpensive algorithm. By analyzing 29 baseball games played in the 2016 and 2017 seasons, we demonstrate that our approach properly reduces the risk of false alarm and generates the results closer to expert-chosen highlights than that of the peak-finding method.