Material Appearance and Perception: Estimating and Generating Properties Across Scales

The rapid advancements in computational imaging and material appearance modeling have opened new opportunities for understanding and simulating the complex interactions between light and materials. This Ph.D. thesis presents a comprehensive exploration of these domains, addressing challenges ranging from stereoscopic image generation to material property estimation and anisotropy measurement.
The first part of the research introduces StereoDiffusion, a novel training-free approach for generating high-quality stereoscopic image pairs using latent diffusion models. This method effectively leverages existing generative models without requiring extensive fine-tuning, enabling fast and reliable stereo image generation. Building on this foundation, the thesis progresses to Materialist, a framework for physically based single-view material estimation and editing. By combining learning-based methods with differentiable rendering, this approach achieves material predictions and physically based edits under diverse lighting conditions, bridging the gap between artistic flexibility and physical realism.
In the latter part of the thesis, the focus shifts to microscopic material properties, specifically anisotropy. A method for Noninvasive Material Anisotropy Estimation is proposed, employing elliptical fitting and machine learning to estimate anisotropy with minimal measurements. Finally, the research introduces a Quantitative Measurement of Turbid Medium Anisotropy technique using polarization imaging, which demonstrates robustness and accuracy in capturing anisotropic properties of translucent materials.
Together, these contributions represent a cohesive advancement in computational imaging and material modeling. By addressing both macroscopic and microscopic challenges, this work not only enhances theoretical understanding but also provides practical tools and methods for a wide range of applications, from extended reality to food science. The research establishes a solid foundation for future exploration and innovation in these dynamic and interdisciplinary fields.