穿插式照明計算中虛擬點光源取樣的改善方法

Abstract

Instant radiosity is an elegant method for rendering global illumination effects by casting virtual point lights (VPL) from primary light sources onto the objects in the scene. The final image is obtained by accumulating the participation of the primary sources and the VPLs. However, it is common to cast tens or hundreds of thousands of VPLs to obtain accurate global illumination, especially in scenes with important occlusion and complex shapes. Interleaved shading consists in reorganizing the image-space into multiple identical size blocks, in which each pixel is shaded using a different subset of VPLs. This method improves rendering speed, breaks hard shadow edges and is trivially parallel; however, the incoherent shading introduces a distracting structured noise. We present here a novel method to improve classic interleaved shading by rearranging the VPL subsets and improving VPL sampling within subsets. Few samples are taken in image space to estimate VPLs visual importance. This serves as a key to sort VPLs and assign them to subsets using a card dealing algorithm. During the rendering phase, a fraction of the VPLs in the subset corresponding to a particular pixel is sampled to shade it. The whole process can easily be abstracted using a global lighting structure, called metalight, and we show that our method leads to dramatic noise variance reduction in the final picture by adding a small fraction of computation. The implementation is straight-forward and can be easily integrated into any interleaved shading-based frameworks and, in general, to most of local or distributed rendering systems on CPU or GPU. We also present different image-space assignment schemes for the VPL subsets to break the regularity of the noise pattern or to adapt it to a simple antialiasing algorithm. Instant radiosity is an elegant method for rendering global illumination effects by casting virtual point lights (VPL) from primary light sources onto the objects in the scene. The final image is obtained by accumulating the participation of the primary sources and the VPLs. However, it is common to cast tens or hundreds of thousands of VPLs to obtain accurate global illumination, especially in scenes with important occlusion and complex shapes. Interleaved shading consists in reorganizing the image-space into multiple identical size blocks, in which each pixel is shaded using a different subset of VPLs. This method improves rendering speed, breaks hard shadow edges and is trivially parallel; however, the incoherent shading introduces a distracting structured noise. We present here a novel method to improve classic interleaved shading by rearranging the VPL subsets and improving VPL sampling within subsets. Few samples are taken in image space to estimate VPLs visual importance. This serves as a key to sort VPLs and assign them to subsets using a card dealing algorithm. During the rendering phase, a fraction of the VPLs in the subset corresponding to a particular pixel is sampled to shade it. The whole process can easily be abstracted using a global lighting structure, called metalight, and we show that our method leads to dramatic noise variance reduction in the final picture by adding a small fraction of computation. The implementation is straight-forward and can be easily integrated into any interleaved shading-based frameworks and, in general, to most of local or distributed rendering systems on CPU or GPU. We also present different image-space assignment schemes for the VPL subsets to break the regularity of the noise pattern or to adapt it to a simple antialiasing algorithm.