Reproduce “ReGenHuman: Re-Generating Human Appearances for Realistic Full-Body Video Anonymization” 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

Adam Sun, Eshaan Barkataki, Arnold Milstein, Gordon Wetzstein, Ehsan Adeli

ReGenHuman anonymizes human-centric video by discarding source human pixels and generating replacement people from pose, depth, and captions with a video diffusion model. StructHuman retains original non-human pixels, while StructAll uses depth for the scene. On specified tests, StructHuman achieves 1.4% Rank-1 re-identification and 85.20% clinical video question-answering accuracy; pose and scene cues limit its privacy guarantee.

Key insights

  • Synthesizing new, identity-free human appearances from structural cues offers a more robust and verifiable architectural privacy guarantee compared to traditional pixel-editing or partial generative methods.
  • Leveraging advanced video-to-video diffusion models with joint temporal denoising is crucial for generating anonymized videos that are both realistic and temporally consistent, overcoming a major limitation of prior generative approaches.
  • A comprehensive, multi-axis evaluation protocol considering privacy, quality, and utility is essential for accurately assessing the performance and trade-offs of full-body video anonymization solutions.

Summary by alphaXiv.

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