Deep intrinsic decomposition trained on surreal scenes yet with realistic light effects

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Intrinsic image estimation is crucial for computer vision tasks but hindered by limited and unrealistic ground-truth datasets.
• Current end-to-end deep learning methods show promise but often neglect essential physical principles.

Purpose of the Study:

• To develop a versatile framework for generating larger, more realistic intrinsic image datasets.
• To create a flexible deep learning architecture that integrates physical constraints via intrinsic losses for improved estimation.

Main Methods:

• A twofold framework combining flexible image generation with a physics-informed deep learning architecture.
• Utilizing intrinsic losses to enforce physical properties within the model.

Main Results:

• The proposed method overcomes classical dataset limitations, offering coherent lighting appearance and larger scale.
• Achieved state-of-the-art results in intrinsic image estimation with low computation time.

Conclusions:

• The presented framework offers a versatile and efficient solution for intrinsic image estimation.
• Integrating physical insights significantly improves deep learning-based intrinsic image estimation performance.