ARTFEED — Contemporary Art Intelligence

FactorizedHMR: Hybrid Framework for Video Human Mesh Recovery

other · 2026-05-16

FactorizedHMR is a two-stage framework for video human mesh recovery that addresses ambiguity in 3D body reconstruction. It uses a deterministic regression module to recover a stable torso-root anchor, then a probabilistic flow-matching module to complete non-torso articulation. The method combines composite target representation, geometry-aware supervision, and feature-aware classifier-free guidance to improve recovery of ambiguous limbs. A synthetic data pipeline supports training. The paper is published on arXiv under ID 2605.14854.

Key facts

  • FactorizedHMR is a two-stage framework for video human mesh recovery.
  • It uses a deterministic regression module for torso-root anchor.
  • A probabilistic flow-matching module handles non-torso articulation.
  • Composite target representation and geometry-aware supervision are used.
  • Feature-aware classifier-free guidance preserves torso-root anchor.
  • A synthetic data pipeline is introduced for training.
  • The paper is on arXiv with ID 2605.14854.
  • The method addresses ambiguity in 3D body reconstruction under occlusion.

Entities

Institutions

  • arXiv

Sources