Object Appearance Representation and Rendering: A Neural Approach to Light Transport Simulation in Translucent media

From the lifelike characters in movies to immersive environments of virtual reality, computer graphics plays a pivotal role in bridging the gap between the real and virtual worlds. At the heart of this pursuit lies the challenge of simulating how light interacts with materials to achieve photorealistic visuals. However, simulating the infinite number of light paths is impractical, necessitating the use of alternative methods.

Monte-Carlo methods are foundational in computer graphics for rendering realistic images by simulating the complex scattering behavior of light. However, these methods are computationally expensive to evaluate since a large number of samples are often necessary for noise-free image synthesis. Naturally occurring organic materials such as skin, fruits, and liquids, characterized by their soft natural look, are important for visual realism. They contain the effects of translucency when the light enters the object following a series of complex internal scattering before exiting the medium from a different location. This makes translucency challenging to both simulate and capture from real-world observations, as compared to the more commonly used surface reflectance models.

In this thesis, we aim to develop efficient and practical methods for representing the appearance of translucent objects, addressing the particularly challenging tasks of simulation and devising a method for real-world capture. Based on our findings, we propose several practical models for representing object appearance that incorporate translucency, spanning applications from interactive and real-time rendering to high-accuracy, assumption-free BSSRDF models designed to enhance representability.

Furthermore, we investigate object capture methods for subsurface scattering, assuming a known object geometry, and evaluate them on synthetic data.