Smooth shadow boundaries with exponentially warped Gaussian filtering

Shadow mapping is widely used in computer graphics for efficiently rendering shadows in real-time applications. Shadow maps cannot be filtered as regular textures, thus their limited resolution can cause severe shadow map discretization artifacts in the rendered images. To solve this problem, several techniques have been proposed, including variance shadow maps (VSM) and exponential shadow maps (ESM). However, these techniques introduce different kinds of “light leaking” artifacts, which are clearly visible in moderately complex scenes. In this paper we propose a new statistical filtering method that approximates the cumulative distribution function (CDF) of depth values by a Gaussian CDF instead of bounding it with Chebyshev Inequality. This approximation significantly reduces “light leaks” and has similar performance and storage requirements compared to the original variance shadow map method. We also show that the combination of this technique with an exponential warp allows us to further reduce the remaining shadowing artifacts from the rendered image.

Figure optionsDownload high-quality image (159 K)Download as PowerPoint slideHighlights
► Our Gaussian CDF and Power CDF approaches are able of improving shadow map filtering and highly remove light leaking artifacts introduced by VSM.
► Performance is maintained relative to VSM and ESM.
► Our technique can be warped by a exponential function to completely remove light leaking artifacts.