UniEgoMotion: A Unified Model for Egocentric Motion Reconstruction, Forecasting, and Generation.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Robotics

Background:

• Egocentric motion understanding is vital for AR/VR, robotics, and healthcare, but current methods struggle with first-person visual data.
• Existing approaches often rely on third-person perspectives and explicit 3D scene data, limiting real-world egocentric applications.

Purpose of the Study:

• To develop a unified framework for egocentric motion generation and forecasting using only first-person images.
• To address the limitations of existing methods in handling egocentric visual data, occlusions, and dynamic cameras.

Main Methods:

• Proposed UniEgoMotion, a conditional motion diffusion model with a head-centric motion representation.
• Utilized first-person images for scene-aware motion synthesis without explicit 3D scene reconstruction.
• Introduced EE4D-Motion, a large-scale dataset for egocentric motion modeling.

Main Results:

• Achieved state-of-the-art performance in egocentric motion reconstruction.
• Demonstrated the first successful generation of human motion from a single egocentric image.
• Validated the effectiveness of the unified framework in extracting scene context for plausible 3D motion inference.

Conclusions:

• UniEgoMotion provides a unified and effective solution for egocentric motion reconstruction, forecasting, and generation.
• The model sets a new benchmark for egocentric motion modeling, opening new avenues for egocentric applications.
• Scene-aware egocentric motion synthesis is achievable using first-person visual inputs.