PBR-Net: Imitating Physically Based Rendering using Deep Neural Network.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Machine Learning

Background:

• Physically based rendering (PBR) generates realistic images for entertainment and synthetic data generation.
• Traditional PBR relies on computationally expensive ray-tracing, limiting its efficiency and parallelization.

Purpose of the Study:

• To develop an efficient, end-to-end deep learning approach for physically based rendering.
• To overcome the computational limitations of traditional ray-tracing methods.

Main Methods:

• Proposed a two-stacked neural network system for efficient PBR.
• The shading network predicts shading images from surface normal, depth, and illumination.
• The composition network combines shading with reflectance for the final image.

Main Results:

• The deep learning approach achieves efficient and photo-realistic image generation.
• The system demonstrates robustness to noise using a modified perceptual loss.
• Outperforms traditional PBR in complex scenes within a reasonable time budget.

Conclusions:

• The proposed deep learning method offers a computationally efficient alternative to traditional PBR.
• The intrinsic image decomposition-inspired approach provides physically plausible results.
• This method is suitable for generating high-fidelity synthetic data and real-time rendering applications.