Reproduce “Human4K: A Large-Scale 4K Multi-View Mocap Dataset for Whole-Body 3D Human Reconstruction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tianshun Han, Ziyu Shi, Lijian Liu, Ajian Liu, Benjia Zhou, Hugo Jair Escalante, Yanyan Liang, Sergio Escalera and 2 more

Human4K is a large-scale dataset comprising millions of 4K multi-view images with mocap-accurate SMPL-X annotations for whole-body 3D human reconstruction, including precise details for hands and faces. This dataset enables models to achieve improved accuracy, especially for extremities and challenging poses, with training on Human4K reducing overall MPVPE on EHF by approximately 11% for Hand4Whole when combined with public datasets.

Key insights

  • Human4K provides complementary high-fidelity supervision, significantly enhancing accuracy for extremities and overall whole-body reconstruction when integrated with existing public datasets.
  • The 4K resolution is crucial for capturing and preserving fine structures, leading to substantial error reductions for hands (e.g., MPVPE from 95.34 mm at 1K to 61.58 mm at 4K) and faces.
  • The dataset's diverse range of actions, extreme articulations, and frequent self-occlusions establishes it as a challenging in-domain benchmark, exposing and addressing limitations of current models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.