Researchers from META* Reality Labs recently unveiled the Hot3D dataset, a groundbreaking collection that is set to accelerate the advancement of computer vision and robotics technologies. The dataset, which was published on the ARXIV platform, features high-quality 3D videos captured from an egocentric perspective, showcasing human interactions with various objects.
Hot3D comprises over 833 minutes of videos (equivalent to about 3.7 million images) filmed using Project ARIA points and the QUEST 3 helmet. The recordings involve 19 individuals engaging with 33 different objects in both residential and office settings. Detailed annotations include 3D representations of objects, hands, and cameras, as well as 3D models of hands and objects.
Experiments have validated the efficacy of the Hot3D dataset in addressing tasks such as 3D hand tracking, object recognition, and object motion modeling. The use of a multi-camera approach in Hot3D has yielded substantially better results compared to single-camera systems.
The dataset is available for use by researchers worldwide and is poised to advance the development of human-machine interfaces, augmented and virtual reality systems, and enhance the interaction of robots with their surroundings.
*META and its products have been deemed extremist, leading to their prohibition within the territory of the Russian Federation.