Generative model-enhanced human motion prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human motion prediction is complex due to action heterogeneity and compositionality.
• Existing models struggle with out-of-distribution (OoD) data, limiting real-world applicability.
• Robustness to distributional shifts is crucial for reliable human motion prediction.

Purpose of the Study:

• To develop a benchmark for evaluating out-of-distribution (OoD) generalization in human motion prediction.
• To introduce a hybrid framework that enhances the robustness of discriminative models to OoD challenges.
• To improve the interpretability of human motion prediction models.

Main Methods:

• Formulated a new OoD benchmark using Human3.6M and CMU motion capture datasets.
• Developed a hybrid framework combining discriminative and generative models.
• Augmented state-of-the-art discriminative architectures with a generative component.

Main Results:

• The proposed hybrid framework significantly improves OoD robustness in human motion prediction.
• In-distribution performance is maintained without sacrifice.
• The approach theoretically facilitates enhanced model interpretability.

Conclusions:

• Human motion predictors must be designed with OoD challenges as a primary consideration.
• The provided framework offers an extensible solution for hardening diverse discriminative architectures against extreme distributional shifts.
• This work advances the development of more reliable and generalizable human motion prediction systems.