Improving human motion generation based on a head-mounted display and its controllers via noise-augmented motion data...

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Existing full-body human motion generation methods using head-mounted displays (HMDs) and controllers assume rigid connections.
• This assumption leads to poor robustness against motion noise common in virtual reality (VR) due to loosely attached devices.

Purpose of the Study:

• To develop a robust human motion generation strategy addressing noise sensitivity.
• To introduce and evaluate the Recurrent Inference Model (RIM) for enhanced full-body motion synthesis.

Main Methods:

• Proposed a generative strategy synthesizing noise-augmented motion data.
• Compared noise-augmentation against direct, fine-tuning, and additive training strategies.
• Developed the Recurrent Inference Model (RIM) for improved motion generation.

Main Results:

• The noise-augmented data strategy enhanced human motion generation model performance.
• The RIM model trained on noise-augmented data showed consistent improvements over SOTA methods.
• Evaluations demonstrated superior performance in both offline and real-time scenarios.

Conclusions:

• Noise-augmented data synthesis is an effective strategy for robust human motion generation.
• The Recurrent Inference Model (RIM) represents a significant advancement in SOTA human motion generation.
• The proposed methods offer improved performance for VR applications requiring accurate motion capture.