Fitted virtual shadow maps

Abstract

Too little shadow map resolution and resulting undersampling artifacts, perspective and projection aliasing, have long been a fundamental problem of shadowing scenes with shadow mapping. We present a new smart, real-time shadow mapping algorithm that virtually increases the resolution of the shadow map beyond the GPU hardware limit where needed. We first sample the scene from the eye-point on the GPU to get the needed shadow map resolution in different parts of the scene. We then process the resulting data on the CPU and finally arrive at a hierarchical grid structure, which we traverse in kd-tree fashion, shadowing the scene with shadow map tiles where needed. Shadow quality can be traded for speed through an intuitive parameter, with a homogenous quality reduction in the whole scene, down to normal shadow mapping. This allows the algorithm to be used on a wide range of hardware.