Learning Pose Controllable Human Reconstruction With Dynamic Implicit Fields From a Single Image.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Generating controllable 3D human models from single images is challenging.
• Existing methods often produce static models, lacking pose adaptability.
• Controlling human model pose is crucial for applications like animation and virtual reality.

Purpose of the Study:

• To develop a method for pose-controllable human reconstruction from a single RGB image.
• To learn a dynamic implicit field capable of representing humans across multiple poses.
• To overcome limitations of static reconstruction and shape-pose entanglement.

Main Methods:

• Constructed a feature-embedded human model (FEHM) for feature propagation across poses.
• Encoded three pose-decoupled features: global image, spatial color, and spatial geometry.
• Designed novel implicit functions to predict dynamic human implicit fields.
• Utilized a large-scale realistic human avatar dataset (SimuSCAN) for supervision.

Main Results:

• Achieved state-of-the-art performance in 3D human reconstruction.
• Demonstrated effective pose control and shape preservation.
• Successfully decoupled human shape from pose information.

Conclusions:

• The proposed method enables high-fidelity, pose-controllable human reconstruction from single images.
• The FEHM and pose-decoupled features are effective for dynamic human modeling.
• This work advances the field of single-image 3D human reconstruction.