Benchmarking Classical and ReSTIR-Based Sampling Techniques for Real-Time Ray Tracing

Low, Jia Ding (2026) Benchmarking Classical and ReSTIR-Based Sampling Techniques for Real-Time Ray Tracing. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text RST_Low Jia Ding_Fulltext.pdf Restricted to Registered users only Download (34MB)

Abstract

Ray tracing is widely used in modern rendering systems due to its ability to produce physically accurate lighting. Recent sampling techniques such as Next Event Estimation (NEE), Light-Source Sampling, and ReSTIR-based Direct and Global Illumination have significantly improved rendering efficiency, yet objective comparison between sampling strategies remains challenging. This project presents a GPU-multithreaded ray tracing benchmarking framework for evaluating both naive and modern sampling techniques under controlled conditions. The system compares Uniform, Cosine-Weighted, GGX, BRDF, Light-Source Sampling, NEE, ReSTIR Direct Illumination, and ReSTIR Global Illumination using a test scene consisting of an enclosed box with multiple plane lights and complex materials. Quantitative evaluation is performed using Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR) while varying sample count and bounce depth. The framework enables objective analysis of convergence behavior and performance, providing a practical tool for research, development, and education.