Something-Something V2 Action Recognition Dataset

Something-Something V2 is a large dataset of annotated video clips of basic actions between humans and some objects in daily life.

This dataset was created by a large number of crowd workers. It allows machine learning models to have a detailed understanding of basic actions that occur in the real world. The dataset contains 220,847 videos, of which 168,913 are training sets, 24,777 are validation sets, and 27,157 are test sets. There are 174 action category annotations.

This is the second version of the dataset. Compared with the previous version, the number of videos has been increased and the image format has been updated from JPG to Webm.