A Novel Adaptive Sampling by Tsallis Entropy

Monte Carlo is the only choice of physically correct method to compute the problem of global illumination in the field of realistic image synthesis. Adaptive sampling is an appealing tool to eliminate noise, which is one of the main problems of Monte Carlo based global illumination algorithms. In this paper, we investigate the use of entropy in the domain of information theory to measure pixel quality and to do adaptive sampling. Especially we explore the nonextensive Tsallis entropy, in which a real number q is introduced as the entropic index that presents the degree of nonextensivity, to evaluate pixel quality. By utilizing the least-squares design, an entropic index q can be obtained systematically to run adaptive sampling effectively. Implementation results show that the Tsallis entropy driven adaptive sampling significantly outperforms the existing methods.