Achieving view-distance and -angle invariance in motion prediction using a simple network.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human motion prediction is crucial for various applications but current models struggle with real-world viewing variations.
• Existing methods often fail due to unaddressed changes in viewing distance and angle during data collection and inference.

Purpose of the Study:

• To develop a human motion prediction model that is invariant to viewing distance and angle.
• To enhance the robustness and applicability of motion prediction in practical, unconstrained environments.

Main Methods:

• Employed Riemannian geometry to constrain neural network learning for invariance.
• Introduced a novel motion space using the path transport square-root velocity function to encode motion sequences.
• Utilized geometric methods for motion coding to linearize optimization and extract motion information.

Main Results:

• Achieved competitive performance in human motion prediction tasks.
• Demonstrated significant invariance to variations in viewing distance and angle.
• Validated the framework on Human 3.6M and CMU MoCap datasets.

Conclusions:

• The proposed Riemannian geometry-based framework enables robust human motion prediction invariant to viewing conditions.
• The method enhances the practical utility of motion prediction models by addressing real-world viewing variations.
• A simple network architecture combined with geometric encoding yields effective and invariant motion prediction.