Reproduce “PeopleSansPeople: A Synthetic Data Generator for Human-Centric Computer Vision” 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

Salehe Erfanian Ebadi, You-Cyuan Jhang, Alex Zook, Saurav Dhakad, Adam Crespi, Pete Parisi, Steven Borkman, Jonathan Hogins and 1 more

PEOPLESANSPEOPLE, developed by Unity Technologies, is a synthetic data generator for human-centric computer vision that leverages the Unity engine and extensive domain randomization to produce diverse, perfectly labeled datasets. Its synthetic pre-training substantially improves model performance in human pose estimation, especially in few-shot learning scenarios, consistently outperforming ImageNet pre-training.

Key insights

  • Synthetic data pre-training provides substantial performance gains for human-centric computer vision tasks, particularly in data-scarce (few-shot) settings.
  • Procedurally generated synthetic data, using techniques like domain randomization, can achieve broader and more diverse human pose distributions than real-world datasets like COCO.
  • The framework effectively bridges the sim2real gap for human pose estimation by providing highly customizable and diverse synthetic training examples.

Summary by alphaXiv.

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