ActivityNet

by Fabian Caba Heilbron,Victor Escorcia,Bernard Ghanem,Juan Carlos NieblesUnknown

ActivityNet

ActivityNet is a new large-scale video benchmark for human activity understanding. ActivityNet aims at covering a wide range of complex human activities that are of interest to people in their daily living. In its current version, ActivityNet provides samples from 203 activity classes with an average of 137 untrimmed videos per class and 1.41 activity instances per video, for a total of 849 video hours. We heavily rely on the crowd and specifically, Amazon Mechanical Turk, to help acquire and annotate ActivityNet. Our acquisition pipeline has three main steps: Collection, Filtering, and Temporal Localization. Collecting candidate videos: We search candidate videos in the web for each ActivityNet category (including expanded queries) Filtering candidate videos: Retrieved videos are verified by Amazon Mechanical Turk workers and clean out those videos which are not related with any intended activity. Temporal localization: Temporal localization provides starting and ending times in which the activity is performed (multiple Amazon Mechanical Turk workers annotate a single video)

Dataset Attributes

Label SVG
TasksEvent Detection
Label SVG
CategoriesGeneral
Label SVG
SensorRGB Camera