OoD

OoD improves robustness of human motion prediction by integrating generative models with discriminative architectures to mitigate out-of-distribution (OoD) failures caused by action heterogeneity and compositionality.

Key Features:

• Hybrid generative–discriminative integration: Integrates generative models with state-of-the-art discriminative architectures for human motion prediction.
• OoD robustness: Enhances out-of-distribution (OoD) robustness in discriminative models for predicting human motion.
• Preserves in-distribution performance: Mitigates OoD failures while maintaining or improving in-distribution predictive performance.
• Action heterogeneity and compositionality: Explicitly addresses heterogeneity and compositionality of actions that induce distributional shifts in predictive modeling.
• Theoretical interpretability: Theoretically enhances interpretability of augmented discriminative architectures.
• Benchmarking datasets: Establishes a benchmark for OoD evaluation using Human3.6M and Carnegie Mellon University (CMU) motion capture datasets.
• Generalizable framework: Formulates a generalizable framework applicable across different discriminative architectures.

Scientific Applications:

• Human motion prediction: Improving robustness and accuracy of models that predict human skeletal motion trajectories.
• OoD evaluation and benchmarking: Providing a benchmark and evaluation protocol for out-of-distribution challenges using Human3.6M and CMU motion capture data.
• Interpretability research: Studying interpretability of discriminative architectures augmented with generative components in motion modeling.

Methodology:

Integrates generative models with discriminative architectures and evaluates OoD robustness and predictive performance on Human3.6M and CMU motion capture datasets.