Hi5: 2D Hand Pose Estimation with Zero Human Annotation

We introduce Hi5, a synthetic hand pose estimation dataset of $\mathbf{583,000}$ images generated without human annotation, leveraging high-fidelity 3D hand models with diverse genders and skin tones, dynamic environments, and consumer-grade rendering. The data synthesis pipeline precisely controls pose diversity and representation, enabling robust model training. Models trained on Hi5 outperform real-data baselines on occluded and perturbed real benchmarks, demonstrating synthetic data’s potential for scalable hand pose estimation.

Recommended citation: Hasan, M., Ozel, C., Long, N., Martin, A., Potter, S., Adnan, T., Lee, S., Zadeh, A. and Hoque, E., 2024. Hi5: 2D Hand Pose Estimation with Zero Human Annotation. arXiv preprint arXiv:2406.03599.
Download Paper