ActivityNet: A large-scale video benchmark for human activity understanding

TLDR

Existing human action datasets are limited in variability and complexity, focusing mainly on simple, manually trimmed actions. This paper introduces ActivityNet, a large‑scale benchmark designed to advance human activity understanding. ActivityNet contains 203 activity classes with an average of 137 untrimmed videos per class (≈849 hours total) and supports untrimmed classification, trimmed classification, and activity detection tasks.

Abstract

In spite of many dataset efforts for human action recognition, current computer vision algorithms are still severely limited in terms of the variability and complexity of the actions that they can recognize. This is in part due to the simplicity of current benchmarks, which mostly focus on simple actions and movements occurring on manually trimmed videos. In this paper we introduce ActivityNet, a new large-scale video benchmark for human activity understanding. Our benchmark 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 illustrate three scenarios in which ActivityNet can be used to compare algorithms for human activity understanding: untrimmed video classification, trimmed activity classification and activity detection.