Reproduce “Human-Art: A Versatile Human-Centric Dataset Bridging Natural and Artificial Scenes” 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

Xuan Ju, Ailing Zeng, Jianan Wang, Qiang Xu, Lei Zhang

A new large-scale dataset, Human-Art, introduces diverse human representations from natural and artificial scenes, providing over 123,000 instances with detailed 2D keypoints and self-contact annotations. This resource enables computer vision models to improve performance across varied art forms, with joint training boosting human detection by an average of 31% and 2D pose estimation by 21% on the dataset's challenging artificial scenes.

Key insights

  • State-of-the-art human-centric computer vision models, when trained solely on natural images, demonstrate a significant domain gap and poor generalization to artificial human representations, often exhibiting over-confidence despite low accuracy.
  • Incorporating diverse artistic and stylized human representations through a dataset like Human-Art is crucial for improving the robustness and generalizability of CV models across various appearances, poses, and textures.
  • Specific annotations such as extended keypoints and self-contact points provide valuable constraints for complex tasks like 3D human mesh recovery, helping to resolve ambiguities and improve the plausibility of reconstructed models, especially in artistic contexts.

Summary by alphaXiv.

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