Nonisotropic Gaussian Diffusion for Realistic 3D Human Motion Prediction

2025-01-10

AI Summary

Existing 3D human motion prediction methods often produce physically implausible artifacts—such as limb stretching and jittering—due to insufficient modeling of biomechanical constraints. To address this, we propose the first latent-space diffusion model explicitly incorporating human kinematic priors. Our approach introduces an anisotropic Gaussian noise schedule and embeds skeletal topology and joint kinematics as explicit inductive biases into both architecture and training objectives. We further design a kinematic consistency regularizer that implicitly enforces skeletal constraints. Evaluated on three real-world datasets, our method significantly improves the physical plausibility and visual realism of predicted motions, achieving state-of-the-art performance. Crucially, ablation studies reveal that widely adopted diversity metrics—e.g., MPJPE-std and FID—exhibit weak correlation with physical plausibility, indicating a fundamental misalignment between current evaluation paradigms and biomechanical fidelity. This finding calls for a rethinking of motion prediction assessment criteria.

Technology Category

• Humans and AI: Human-Aware Planning and Behavior Prediction
• Computer Vision: Motion & Tracking
• Cognitive Modeling & Cognitive Systems: Simulating Human Behavior

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Probabilistic human motion prediction aims to forecast multiple possible future movements from past observations. While current approaches report high diversity and realism, they often generate motions with undetected limb stretching and jitter. To address this, we introduce SkeletonDiffusion, a latent diffusion model that embeds an explicit inductive bias on the human body within its architecture and training. Our model is trained with a novel nonisotropic Gaussian diffusion formulation that aligns with the natural kinematic structure of the human skeleton. Results show that our approach outperforms conventional isotropic alternatives, consistently generating realistic predictions while avoiding artifacts such as limb distortion. Additionally, we identify a limitation in commonly used diversity metrics, which may inadvertently favor models that produce inconsistent limb lengths within the same sequence. SkeletonDiffusion sets a new benchmark on three real-world datasets, outperforming various baselines across multiple evaluation metrics. Visit our project page: https://ceveloper.github.io/publications/skeletondiffusion/

Problem

• 3D human motion prediction
• limb length anomaly
• unnatural motion

Innovation

• SkeletonDiffusion
• 3D Human Motion Prediction
• Anatomical Constraints