H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Huang, Zhanbo, Liu, Xiaoming, Kong, Yu
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911704927961088
author Huang, Zhanbo
Liu, Xiaoming
Kong, Yu
author_facet Huang, Zhanbo
Liu, Xiaoming
Kong, Yu
contents Parametric human models capture global pose but cannot represent the non-rigid surface dynamics of clothing and soft tissue. Generic scene flow estimates dense motion but breaks down on articulated bodies, where pixel-level supervision is also intractable to acquire. We introduce H-Flow, a dense human scene flow that captures both skeletal kinematics and surface deformation. A unified multi-head transformer estimates flow from monocular video, jointly predicting pose and depth as companion outputs. The challenge lies in the lack of supervision. In place of unattainable labels, we anchor the network in the physics of human motion, encoding geometric, structural, and biomechanical priors as cross-modal training objectives. We further introduce DynAct4D, a high-fidelity synthetic benchmark providing dense flow annotations across diverse subjects, garments, and motions. On standard benchmarks, H-Flow outperforms scene-flow and parametric baselines, and generalizes zero-shot to in-the-wild video. Code, models, and the DynAct4D benchmark will be released upon publication
format Preprint
id arxiv_https___arxiv_org_abs_2605_22629
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning
Huang, Zhanbo
Liu, Xiaoming
Kong, Yu
Computer Vision and Pattern Recognition
Parametric human models capture global pose but cannot represent the non-rigid surface dynamics of clothing and soft tissue. Generic scene flow estimates dense motion but breaks down on articulated bodies, where pixel-level supervision is also intractable to acquire. We introduce H-Flow, a dense human scene flow that captures both skeletal kinematics and surface deformation. A unified multi-head transformer estimates flow from monocular video, jointly predicting pose and depth as companion outputs. The challenge lies in the lack of supervision. In place of unattainable labels, we anchor the network in the physics of human motion, encoding geometric, structural, and biomechanical priors as cross-modal training objectives. We further introduce DynAct4D, a high-fidelity synthetic benchmark providing dense flow annotations across diverse subjects, garments, and motions. On standard benchmarks, H-Flow outperforms scene-flow and parametric baselines, and generalizes zero-shot to in-the-wild video. Code, models, and the DynAct4D benchmark will be released upon publication
title H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2605.22629