Intrinsic Image Decomposition Using Paradigms.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Intrinsic image decomposition separates scene reflectance (albedo) and illumination (shading).
• Current methods often rely on supervised learning with rendered data or human judgments, which can be inconvenient or limiting.
• Existing approaches struggle to explain how humans learn intrinsic images without explicit models.

Purpose of the Study:

• To develop an unsupervised method for intrinsic image decomposition.
• To learn albedo and shading without human annotations, rendered data, or ground truth.
• To provide a method that aligns with how humans might learn intrinsic properties.

Main Methods:

• Utilizes spatial models for albedo and shading.
• Trains a neural network on synthetic images created by multiplying sampled albedo and shading fields.
• Employs a novel smoothing procedure for short-scale accuracy and an averaging procedure for long-scale error control and equivariance.

Main Results:

• Achieves Wasserstein-Hadamard distance (WHDR) scores competitive with supervised methods.
• Produces high-quality albedo and shading maps, preserving fine details like wood grain and grooves.
• Demonstrates the potential for unsupervised learning in intrinsic image decomposition.

Conclusions:

• The proposed unsupervised method is effective for intrinsic image decomposition.
• The approach offers a viable alternative to supervised methods, reducing reliance on extensive datasets.
• The large test/train variance in WHDR scores suggests caution when interpreting small performance differences.