Investigating Pose Representations and Motion Contexts Modeling for 3D Motion Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Predicting human motion from pose sequences is vital for human-robot interaction.
• Existing methods often overlook the impact of skeletal pose representation on prediction accuracy.
• Recurrent Neural Network (RNN) approaches struggle with long-term dependencies in sequential data.

Purpose of the Study:

• To investigate the effect of various pose representations on motion prediction.
• To propose a novel RNN architecture, AHMR, for enhanced motion prediction.
• To explore geometrically significant loss functions for motion prediction.

Main Methods:

• Conducted an in-depth study on different skeletal pose representation schemes.
• Developed a novel Attentive Hierarchical Motion Recurrent network (AHMR) architecture.
• Utilized geodesic and forward kinematics losses, alongside traditional L2 loss.

Main Results:

• The proposed AHMR model outperforms state-of-the-art methods in short-term motion prediction.
• AHMR demonstrates significantly improved long-term prediction capabilities, maintaining natural motion over extended periods.
• The method shows effectiveness across diverse articulated objects, including humans, fish, and mice.

Conclusions:

• Skeletal pose representation significantly impacts motion prediction performance.
• The AHMR architecture effectively captures both local and global motion contexts.
• Novel geometric loss functions enhance the quality and naturalness of predicted motions.