Meta unveils HOT3D dataset for advanced computer vision training

datasetcomputer-visionhand-trackingegocentric-videoroboticsar-vr

Abstraction: Egocentric 3D hand-object interaction dataset from Meta Reality Labs

Key points:

HOT3D contains 833+ minutes (3.7M+ images) of egocentric multi-view RGB and monochrome video of hand-object interactions
Dataset covers 19 subjects interacting with 33 diverse rigid objects in office and household scenarios
Collected using Meta's Project Aria prototype AR glasses and commercially available Quest 3 VR headset
Ground-truth 3D poses obtained via professional motion-capture with optical markers; hand annotations in UmeTrack and MANO formats
Multi-view models trained on HOT3D significantly outperform single-view counterparts on 3D hand tracking, 6DoF object pose estimation, and 3D lifting
Dataset is publicly available and targets robotics, AR/VR, and human-machine interface research

Connections: Meta · Meta Reality Labs · Computer Vision · Datasets · Egocentric Video · Robot Manipulation