GVHMR: Gravity-View Coordinates for Global Human Motion Recovery From Monocular Videos.

Area of Science:

• Computer Vision
• Human Motion Analysis
• Robotics

Background:

• Recovering accurate 3D human motion from monocular video is challenging due to the inherent ambiguity of world coordinate systems.
• Prior auto-regressive methods suffer from error accumulation, limiting their real-world applicability.
• Existing techniques struggle with precise world-grounding and can produce unnatural motion artifacts.

Purpose of the Study:

• To develop a novel method for robust and accurate world-grounded human motion recovery from monocular video.
• To address the coordinate system ambiguity and error accumulation issues present in previous approaches.
• To improve the realism and stability of recovered human motion, particularly in challenging in-the-wild scenarios.

Main Methods:

• Propose a novel Gravity-View (GV) coordinate system defined by world gravity and camera view direction for per-frame pose estimation.
• Utilize camera rotations to transform estimated poses from GV space back to the world coordinate system, creating global motion sequences.
• Introduce a Stationary Label Predictor (SLP) module to identify stationary states of hands and feet, mitigating artifacts like foot-sliding.

Main Results:

• The proposed GV method significantly reduces ambiguity in image-pose mapping, leading to more accurate pose estimation.
• Per-frame estimation in GV space successfully avoids auto-regressive error accumulation, outperforming state-of-the-art methods.
• Experiments demonstrate superior performance in both camera-space and world-grounded motion recovery, with notable improvements in accuracy and speed.
• The SLP module effectively reduces unnatural motion artifacts and enhances performance with fast-motion inputs.

Conclusions:

• The novel Gravity-View coordinate system offers a robust solution for world-grounded human motion recovery from monocular video.
• The per-frame estimation strategy combined with SLP significantly enhances motion realism and stability, outperforming existing methods.
• This work provides a more accurate and efficient approach for analyzing human motion in unconstrained environments.