*NEW* Our #ECCV2026 @eccvconf paper Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation Now on ArXiv w Dataset, Code&model https://sid2697.github.io/epic-contact/ https://arxiv.org/abs/2606.30598 Two contributions: 1. EPIC-Contact Dataset 2. HOPformer Method &Checkpoint 🧵 1/6
No Digg Deeper questions have been answered for this story yet.
TL;DR we predict bi-manual hand & object pose in a single forward pass from in-the-wild ego frames. Achieved through manually-collected high-quality annotations of 2.3K video clips (62.3K frames) of functional interactions during stable grasps with objects from 9 classes 2/6
*NEW* Our #ECCV2026 @eccvconf paper Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation Now on ArXiv w Dataset, Code&model https://sid2697.github.io/epic-contact/ https://arxiv.org/abs/2606.30598 Two contributions: 1. EPIC-Contact Dataset 2. HOPformer Method &Checkpoint 🧵 1/6
We release EPIC-Contact dataset - manually annotated in-contact hand vertices AND bijective annotations of corresponding object vertices. We contribute a full pipeline from hand painting to efficiently annotate objects, then fit posed hand&object meshes. Manually checked 3/6
TL;DR we predict bi-manual hand & object pose in a single forward pass from in-the-wild ego frames. Achieved through manually-collected high-quality annotations of 2.3K video clips (62.3K frames) of functional interactions during stable grasps with objects from 9 classes 2/6
+ We propose HOPformer (SOTA on Arctic) & eval on EPIC-Contact, feed-forward model that utilises hand prior to cross-attend to object pose backbone & decoder that estimates both poses of hands & object (relative to canonical). + generalisable to new instances of known classes 4/6
Work led by @Sid__Bansal w @zhifan_zhu @terry_jhzhao00 @sha2nk_t and @Michael_J_Black - a 16-months intense collaboration collaboration bw @bristolcs and @MPI_IS. Now @eccvconf #ECCV2026 [don't let anyone convince you good research can be done in a few months!] 5/6
+ We propose HOPformer (SOTA on Arctic) & eval on EPIC-Contact, feed-forward model that utilises hand prior to cross-attend to object pose backbone & decoder that estimates both poses of hands & object (relative to canonical). + generalisable to new instances of known classes 4/6
@Sid__Bansal @zhifan_zhu @terry_jhzhao00 @sha2nk_t @Michael_J_Black @bristolcs @MPI_IS @eccvconf Explore the dataset now at: https://sid2697.github.io/epic-contact/#explorer As we always do: Dataset, Code & models are public already along w ArXiv and full details to reproduce in code and supplementary EPIC-Cotact Dataset: https://huggingface.co/datasets/Sid2697/epic-contact HOPformer Code: https://github.com/Sid2697/HOPformer 6/6
Work led by @Sid__Bansal w @zhifan_zhu @terry_jhzhao00 @sha2nk_t and @Michael_J_Black - a 16-months intense collaboration collaboration bw @bristolcs and @MPI_IS. Now @eccvconf #ECCV2026 [don't let anyone convince you good research can be done in a few months!] 5/6
