Dottorato in INGEGNERIA INFORMATICA

Titolo della tesi: Optimizing Neural Path Guiding with Parametric Mixture Models: A Comprehensive Evaluation and Refinement

This dissertation presents my work during my doctoral research in Computer Graphics. In this work, we explore the domain of realistic rendering, focusing on the development of a physically based GPU path tracer, Vortex, and the innovative technique of Neural Path Guiding, which culminated in the work "Optimizing Neural Path Guiding with Parametric Mixture Models: A Comprehensive Evaluation and Refinement", which has been submitted to "Transaction on Visualization and Computer Graphics" and is currently under review for publication. Vortex was designed from the ground up during my doctoral research to facilitate both the exploration of advanced rendering techniques with neural networks and the development of a comprehensive understanding of the theoretical and practical aspects of rendering and machine learning. The architecture of Vortex is optimized to harness the capabilities of modern hardware, prioritizing near-real-time rendering and efficient scene management. By incorporating the Wavefront Architecture and utilizing CUDA kernels alongside NVIDIA OptiX for accelerated ray tracing, the framework seamlessly integrates machine learning algorithms, such as Neural Path Guiding, to enhance rendering accuracy and optimize light transport simulations. In parallel, the research on Neural Path Guiding led to the work titled "Optimizing Neural Path Guiding with Parametric Mixture Models," which deeply investigates using Neural Path Guiding with Parametric Mixture Models (PMMs). This work does not propose a radically new concept but refines and evaluates existing state-of-the-art methods, achieving new performance benchmarks in Neural Path Guiding with PMMs. By combining Next Event Estimation with Neural Path Guiding and employing advanced input encodings such as Hash Grids and Spherical Harmonics alongside an improved and corrected Normalized Anisotropic Spherical Gaussian (NASG) distribution, the method achieves an accurate representation of light distribution in complex scenes. The integration of Neural Path Guiding with the Wavefront Architecture enables adaptive sampling of light paths guided by neural networks without requiring pre-training. An extensive ablation study further clarifies the influence of various hyperparameters, offering valuable insights into the behavior of Neural Path Guiding techniques. The results demonstrate notable improvements over current state-of-the-art methods, highlighting the potential of this research to advance the field of efficient and realistic rendering.