This paper introduces a video dataset of spatio-temporally localized Atomic Visual Actions (AVA). The AVA dataset densely annotates 80 atomic visual actions in 57.6k movie clips with actions localized in space and time, resulting in 210k action labels with multiple labels per human occurring frequently. The main differences with existing video datasets are: the definition of atomic visual actions, which avoids collecting data for each and every complex action; precise spatio-temporal annotations with possibly multiple annotations for each human; the use of diverse, realistic video material (movies). This departs from existing datasets for spatio-temporal action recognition, such as JHMDB and UCF datasets, which provide annotations for at most 24 composite actions, such as basketball dunk, captured in specific environments, i.e., basketball court.
We implement a state-of-the-art approach for action localization. Despite this, the performance on our dataset remains low and underscores the need for developing new approaches for video understanding. The AVA dataset is the first step in this direction, and enables the measurement of performance and progress in realistic scenarios.