EPIC-Contact dataset and HOPformer bring 3D hand-object pose estimation into the wild

Michael_J_Black · x · 2026-09-10

University of Bristol and MPI researchers presented at ECCV 'Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation': EPIC-Contact, an in-the-wild egocentric dataset (2.3K clips, 62.3K frames from EPIC-Kitchens) with dense bijective 3D hand-object contact annotations enabling supervision without motion capture; and HOPformer, a transformer conditioning object pose on strong hand priors to predict hands and object in a single forward pass under clutter and occlusion. Dataset, code, and checkpoints are released on arXiv/HF.

Original post →

More from Embodied

Embodied channel →