Massively parallel inverse rendering using Multi-objective Particle Swarm Optimization

Koki Nagano, Thomas Joseph Collins, Chi-An Chen, Aiichiro Nakano

Abstract: We present a novel GPU-accelerated per-pixel inverse rendering optimization algorithm based on Particle Swarm Optimization (PSO). Our algorithm estimates the per-pixel scene attributes—including reflectance properties—of a 3D model, and is fast enough to do in situ visualization of the optimization in real-time. The algorithm’s high parallel efficiency is demonstrated through our GPU/GLSL shader implementation of the method. IRPSO is validated experimentally on simulated ground truth images, while a suite of tests performed on the University of Southern California’s High Performance Computing Center cluster provides strong evidence that our method can scale to larger, more difficult inverse rendering problems. Graphical abstract