Moments in Time is a large-scale dynamic moment dataset for identifying and understanding actions in videos. The dataset contains 1 million annotated videos. Each video is 3 seconds long. The subjects of the actions include people, animals, objects and even natural phenomena, and the actions are captured according to the dynamic scene at that time. The intra-class and inter-class differences of the dataset are very large, and there are actions that partially or completely rely on sound information, such as clapping.