MILO: Multi-Bounce Inverse Rendering for Indoor Scene With Light-Emitting Objects.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Recent advances in inverse rendering rely on high-dimensional lighting representations and differentiable rendering.
• Existing methods struggle with complex multi-bounce lighting effects and exhibit light source model deviations and ambiguities, limiting scene editing applications.

Purpose of the Study:

• To develop a multi-bounce inverse rendering method capable of correctly rendering complex lighting effects for scene editing.
• To address limitations in handling multi-bounce lighting and ambiguities in current inverse rendering techniques.

Main Methods:

• Utilized Monte Carlo path tracing for multi-bounce inverse rendering.
• Proposed a novel light source model tailored for indoor scene editing.
• Designed a specific neural network with disambiguation constraints to mitigate inverse rendering ambiguities.

Main Results:

• The method successfully enables correct rendering of complex multi-bounce lighting effects in scene editing.
• Evaluations on synthetic and real indoor scenes demonstrated improved photo-realistic quality through tasks like virtual object insertion, material editing, and relighting.

Conclusions:

• The proposed multi-bounce inverse rendering method enhances scene editing capabilities by accurately capturing complex lighting.
• The novel light source model and neural network design effectively address prior limitations, leading to superior photo-realistic results.