Relightable Detailed Human Reconstruction From Sparse Flashlight Images.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Reconstructing detailed 3D human geometry and appearance from images is challenging.
• Existing methods struggle with accurately capturing surface details and material properties, especially from sparse data.

Purpose of the Study:

• To develop a lightweight system for high-fidelity 3D human reconstruction from sparse flashlight images.
• To enable direct export of reconstructed models for rendering and relighting in traditional graphics pipelines.

Main Methods:

• Utilizes multi-view flashlight images captured with consumer cameras and co-located LEDs.
• Integrates a pre-trained model within a differentiable physics-based rendering framework.
• Jointly optimizes an image feature extractor with MLPs for Signed Distance Function (SDF) and Bidirectional Reflectance Distribution Function (BRDF) prediction.

Main Results:

• Produces detailed geometry, including garment wrinkles and accurate surface reflectance.
• Achieves high-fidelity 3D human models with improved geometry and appearance compared to existing methods.
• Demonstrates effective disambiguation between shading and material properties using view-specific shading cues.

Conclusions:

• The proposed system offers a robust solution for relightable 3D human reconstruction from sparse data.
• Joint optimization within a differentiable rendering framework is key to high-quality reconstruction of real-world details.
• The system provides exportable, detailed 3D human models suitable for advanced graphics applications.