NIID-Net: Adapting Surface Normal Knowledge for Intrinsic Image Decomposition in Indoor Scenes

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Intrinsic image decomposition is crucial for augmented reality (AR) applications, enabling better integration of virtual elements with real scenes.
• The process is ill-posed, particularly in indoor settings with complex lighting and limited real-world training data.

Purpose of the Study:

• To propose NIID-Net, a novel learning-based framework to address the challenges in intrinsic image decomposition.
• To leverage surface normal knowledge for improved reflectance and shading estimation in complex indoor scenes.

Main Methods:

• NIID-Net integrates surface normal estimation knowledge into intrinsic image decomposition.
• Normal feature adapters incorporate scene geometry features.
• An integrated lighting map propagates object contour and planarity information, handling spatially-varying indoor lighting.

Main Results:

• NIID-Net demonstrates competitive performance in reflectance estimation.
• The method significantly outperforms existing approaches in shading estimation, both quantitatively and qualitatively.
• The framework effectively handles complex indoor lighting conditions.

Conclusions:

• NIID-Net offers a robust solution for intrinsic image decomposition, particularly in challenging indoor environments.
• The integration of surface normal knowledge provides a significant advancement for AR applications.
• The proposed methods enhance visual coherence in AR by improving image decomposition accuracy.