A large-scale dataset for recognizing and
understanding action in videos
Moments is a research project dedicated to building a very large-scale dataset to help AI systems recognize and understand actions and events in videos.
Today, the dataset includes a collection of one million labeled 3 second videos, involving people, animals, objects or natural phenomena, that capture the gist of a dynamic scene.
Three seconds events capture an ecosystem of changes in the world: 3 seconds convey meaningful information to understand how agents (human, animal, artificial or natural) transform from one state to another.
Designed to have large inter-class and intra-class variation that represent dynamical events at different levels of abstraction (i.e. "opening" doors, drawers, curtains, presents, eyes, mouths, and even flower petals).
A large-scale, human-annotated video dataset capturing visual and/or audible actions, produced by humans, animals, objects or nature that together allow for the creation of compound activities occurring at longer time scales.
Supervised tasks on a large coverage of the visual and auditory ecosystem help construct powerful but flexible feature detectors, allowing models to quickly transfer learned representations to novel domains.
Can we understand what models attend to during a prediction?
Here, we show the areas of the video frames that our neural network is focusing on in order to recognize the event in the video. These methods show the networks ability to locate the most important areas to focus on for each video clip so that it can identify each moment.
Mathew Monfort, Alex Andonian, Bolei Zhou,
Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown,
Quanfu Fan, Dan Gutfreund, Carl Vondrick, Aude Oliva
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019
Mathew Monfort, Bowen Pan, Kandan Ramakrishnan,
Alex Andonian, Barry A McNamara, Alex Lascelles,
Quanfu Fan, Dan Gutfreund, Rogerio Feris, Aude Oliva
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021
Mathew Monfort*, SouYoung Jin*, Alexander Liu, David Harwath,
Rogerio Feris, James Glass, Aude Oliva
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021
Convolutional neural networks (CNNs) trained on the Moments in Time Dataset can be downloaded and used for action and event recognition in videos.