Abstract:Generating realistic human motion is essential for many computer vision and graphics applications. The wide variety of human body shapes and sizes greatly impacts how people move. However, most existing motion models ignore these differences, relying on a standardized, average body. This leads to uniform motion across different body types, where movements don't match their physical characteristics, limiting diversity. To solve this, we introduce a new approach to develop a generative motion model based on body shape. We show that it's possible to train this model using unpaired data by applying cycle consistency, intuitive physics, and stability constraints, which capture the relationship between identity and movement. The resulting model generates diverse, physically plausible, and dynamically stable human motions that are both quantitatively and qualitatively more realistic than current state-of-the-art methods. More details are available on our project page <a class="link-external link-https" href="https://CarstenEpic.github.io/humos/" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **Generate realistic and physically reasonable dynamically - stable human motions that conform to the differences in human body shapes**. ### Problem Background Existing human motion models usually overlook the diversity of human body shapes and proportions and use standardized average body shapes to generate motions. This results in the homogenization of motion patterns when people with different body shapes perform the same actions, unable to accurately reflect individual body characteristics, thus limiting the diversity and authenticity of motions. ### Paper Goals To solve this problem, the author proposes a new method named **HUMOS**, aiming to learn a body - shape - conditioned human motion generation model. Specifically, the goals of the HUMOS model are: 1. **Generate realistic human motions**: Ensure that the generated motions are visually realistic. 2. **Physical reasonableness**: Ensure that the generated motions conform to physical laws and avoid unreasonable phenomena such as foot - sliding and ground - penetration. 3. **Dynamical stability**: Ensure that the generated motions remain stable during the dynamic process and do not experience imbalance or instability. ### Main Contributions 1. **Identity - preserving cycle - consistency loss**: By introducing cycle - consistency loss, ensure that the model can preserve identity characteristics when regenerating motions. 2. **Differentiable dynamical stability and physical constraints**: Design new dynamical stability and physical constraint terms to ensure that the generated motions not only look realistic but also conform to biomechanical principles. 3. **Self - supervised training with unpaired data**: Utilize the self - supervised learning strategy to solve the problem of lacking paired body - shape and motion data, enabling the model to be effectively trained without paired data. ### Application Scenarios A key application scenario of the HUMOS model is animation retargeting, that is, migrating the actions of one character to another character with a different body shape. Traditional retargeting methods often rely on simple heuristic rules, overlook body - shape differences, and are prone to failure in extreme postures or complex actions. However, HUMOS can effectively learn how people with different body shapes perform the same actions, thereby generating more natural and realistic motions. ### Summary The core problem of the paper is to solve the problem that existing motion models overlook body - shape differences. By introducing a body - shape - conditioned generation model, ensure that the generated motions are both realistic and physically reasonable and can adapt to human bodies with different body shapes.

HUMOS: Human Motion Model Conditioned on Body Shape

Ergonomics-inspired Reshaping and Exploration of Collections of Models

HuMoR: 3D Human Motion Model for Robust Pose Estimation

Shape Conditioned Human Motion Generation with Diffusion Model

Physics-based Human Motion Estimation and Synthesis from Videos

imGHUM: Implicit Generative Models of 3D Human Shape and Articulated Pose

LatentHuman: Shape-and-Pose Disentangled Latent Representation for Human Bodies

Dynamic Motion Transition: A Hybrid Data-driven and Model-driven Method for Human Pose Transitions

Unimotion: Unifying 3D Human Motion Synthesis and Understanding

TEMOS: Generating diverse human motions from textual descriptions

General Automatic Human Shape and Motion Capture Using Volumetric Contour Cues

Champ: Controllable and Consistent Human Image Animation with 3D Parametric Guidance

HuManiFlow: Ancestor-Conditioned Normalising Flows on SO(3) Manifolds for Human Pose and Shape Distribution Estimation

DiverseMotion: Towards Diverse Human Motion Generation Via Discrete Diffusion

Towards Accurate Markerless Human Shape and Pose Estimation over Time

The Wanderings of Odysseus in 3D Scenes

Holistic-Motion2D: Scalable Whole-body Human Motion Generation in 2D Space

Human Motion Generation with Multifactor Models

HUMBI: A Large Multiview Dataset of Human Body Expressions and Benchmark Challenge

HSPACE: Synthetic Parametric Humans Animated in Complex Environments

Action2Motion: Conditioned Generation of 3D Human Motions