The S-Hock dataset: A new benchmark for spectator crowd analysis

Although crowd analysis is a classical and extensively studied problem for the computer vision community, the vast majority of the works in the literature assume a single type of crowd, while the sociological literature classifies a number of different typologies, each one with their own distinctive traits. In this paper we focus on a particular kind of crowd referred in sociology as spectator crowd , which consists a number of people that are “interested in watching something specific that they came to see” Berlonghi (1995). This is the typical social formation that attends entertainment events like sport matches, concerts, movies, etc. In this work we present a novel dataset, the Spectators Hockey ( S-Hock ), containing almost 30 hours of videos recorded at an ice hockey rink during the Winter Universiade “Trentino2013”. On these data we provide a massive annotation, focusing on the spectators at different levels of detail: from high level features describing which team a person supports and if he/she knows his/her neighbors; to a lower level, where we consider standard pose information as well as atomic actions like applauding, jumping, etc. We also provide annotations for the game field, which allows us to analyze the relationship between the crowd behavior and the events of the match. Eventually we provide more than 100 million of annotations, that can be used for many different tasks spanning from standard applications, like people counting and head pose estimation, to higher level tasks, like excitement estimation and automatic summarization. We provide protocols and baseline results for all of these applications, encouraging further research in these field. © 2017 Elsevier Inc. All rights reserved.