Image Coherence Based Adaptive Sampling for Image Synthesis

Due to generality, simplicity and robustness of Monte Carlo, as well as the high complexity of the computation of global illumination problem, Monte Carlo is a very good choice for synthesizing image accounting for global illumination effects. However, the well-known problem in Monte Carlo based methods for global illumination is noise. We explore adaptive sampling as a method to reduce noise. We introduce a coherence distance map, which is one kind of formulization for image coherence, to conduct the adaptive sampling scheme. Based on the coherence distance map, we construct an elegant probability density function to drive Monte Carlo importance sampling to adaptively controlling the number of required samples per pixel. The proposed algorithm can not only improve image quality efficiently, but also be implemented easily. In addition, our approach is unbiased and thus superior to mostly earlier adaptive sampling techniques.