Robust and Accurate 3D Self-Portraits in Seconds.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Computer Graphics

Background:

• Generating high-fidelity 3D self-portraits from single-view depth data presents significant challenges.
• Existing methods struggle with robustness, efficiency, and handling complex clothing.
• Accurate reconstruction of non-rigid human bodies remains an open problem.

Purpose of the Study:

• To develop an efficient and robust method for generating accurate 3D self-portraits using a single RGBD camera.
• To address the limitations of current 3D reconstruction techniques, particularly for subjects in loose clothing.
• To provide a comprehensive benchmark dataset for evaluating single-view 3D self-portrait reconstruction.

Main Methods:

• PIFusion: Combines learning-based 3D recovery with volumetric non-rigid fusion for accurate sparse partial scans.
• Non-rigid volumetric deformation: Continuously refines the learned shape prior for improved accuracy.
• Lightweight bundle adjustment: Ensures consistency and loop closure among partial scans.
• Non-rigid texture optimization: Enhances the realism and quality of the final 3D portrait.

Main Results:

• The proposed method generates detailed and realistic 3D self-portraits in seconds.
• Demonstrates superior ability to reconstruct subjects wearing extremely loose clothing.
• Achieves state-of-the-art performance in accuracy, efficiency, and generality compared to existing methods.
• The new benchmark dataset facilitates rigorous evaluation of single-view 3D reconstruction.

Conclusions:

• PIFusion offers a significant advancement in single-view 3D self-portrait reconstruction.
• The method's efficiency and robustness make it suitable for real-time applications.
• The contributions advance the field of 3D human body modeling and capture.