NIID-Net

NIID-Net performs intrinsic image decomposition of indoor scenes by separating images into reflectance and shading components using learned surface normal knowledge to improve estimation under spatially-varying indoor lighting.

Key Features:

• Intrinsic decomposition: Separates a natural image into reflectance and shading components for indoor scenes.
• Surface normal integration: Leverages surface normal knowledge to inform the decomposition and compensate for scarce intrinsic decomposition labels.
• Normal Feature Adapters: Incorporates geometric scene information derived from surface normals to aid reflectance and shading separation.
• Integrated Lighting Map: Propagates object contour and planarity information during shading rendering to represent spatially-varying indoor lighting.
• Learning-based framework: Uses a learning-based approach that exploits more abundant surface normal data to address the ill-posed nature of intrinsic decomposition.
• Shading performance: Demonstrates significant improvement in shading estimation and competitive reflectance estimation compared to prior methods, reported both quantitatively and qualitatively.

Scientific Applications:

• Augmented reality: Enables visual coherence between virtual elements and real indoor scenes by improving shading estimation.
• Indoor scene analysis: Supports analysis of scenes with complex, spatially-varying indoor lighting.
• Computer vision research: Provides a geometry-aware approach for advancing intrinsic image decomposition and illumination estimation methods.

Methodology:

Employs a learning-based framework that integrates surface normal knowledge via Normal Feature Adapters and an Integrated Lighting Map which propagates object contour and planarity information during shading rendering to represent spatially-varying indoor lighting.