Aff-Wild2 is an extension of the Aff-Wild dataset, mainly used for emotion recognition. This dataset doubles the number of video frames and subjects included, with a total of 558 videos and approximately 2.8 million frames, thereby increasing the variability of the included behaviors and participants.

Currently, AffWild2 is the first large-scale dataset that contains annotations for all three major behavioral tasks, and it is also the first audio-visual dataset with AUs annotations. All AU-annotated datasets do not contain audio, only images or videos.