Overview
EgoHand is a monocular 3D hand reconstruction model that achieves state-of-the-art performance on egocentric video. The work targets a fast-growing area of computer vision in 2026, as first-person wearable cameras (smart glasses, action cameras, and head-mounted AR devices) generate massive volumes of egocentric footage that demand robust, real-time hand tracking.
Training Data
Central to the approach is the curation of a high-quality training corpus of 14 million images, combining:
- Synthetic data
- In-the-wild video
- Multi-camera motion capture
The mixture is designed to balance scene diversity against supervision quality, an increasingly important consideration as 2026 pipelines blend large-scale synthetic generation with real-world capture.
Model Architecture
The model adopts SOMA-Hand as its parametric representation and processes 2D hand crops with a transformer encoder. Two design elements strengthen geometric fidelity:
- Camera cues to improve geometric consistency
- An auxiliary 2D branch to enhance image alignment
Evaluation Results
Zero-shot evaluations on HOT3D, ARCTIC, and H2O demonstrate strong out-of-domain generalization. Compared with competing baselines, EgoHand reports:
- Lower joint-position errors
- Lower root-translation errors
- Lower acceleration errors
All at an inference latency of approximately 9.69 ms, which is well within the real-time budgets expected of 2026 on-device AR and wearable applications.
A blinded user preference study further shows that participants favor EgoHand's reconstructed hand poses over those from Apple EgoDex and Ropedia Xperience-10M.
Availability
The EgoHand model and source code are released under the OpenMDW license, allowing free commercial use of the model.
Publication Details
- arXiv: 2610.30293 [cs.CV]
- Submitted: 21 October 2026
- Subjects: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
- Authors: Chao-Yeh Chen, Tiffany Chen, Lin Duan, Tianyi Hu, Yu-Hsiang Huang, Umar Iqbal, Jiefeng Li, Edy Lim, Yanchen Liu, Yihan Pang, Hang Qiu, Jun Saito, Shao Su, Qingsi Wang, Xiufeng Xie
- Note: Authors are listed alphabetically. Project leads are Chao-Yeh Chen, Tiffany Chen, and Umar Iqbal.
via ArXiv CV