Light Space Perspective Shadow Maps

https://www.cg.tuwien.ac.at/research/publications/2004/Wimmer-2004-LSPM/Wimmer-2004-LSPM-Paper.pdf

abstract
in this paper, we present a new shaow mapping technique that improves upon the quality of persepective and uniform shadow maps. our technique uses a perspective transform specified in light space which allows treating all light as directional lights and does not change the direction of the light sources. this gives all the benefits of the perspective mapping but avoids the problems inherent in perspective shadow mapping like singularies in post-perspective space, missed shadow casters etc. furthermore, we show that both uniform and perspective shadow maps distribute the perspective aliasing error that occurs in shadow mapping unequally over the available depth range.
we therefore propose a transform that equalizes this error and gives equally pleasing results for near and far viewing distances. our method is simple to implement, requires no scene analysis and is therefore as fast as uniform shadow mapping.

1. introduction

shadows belong to the most important effects to convey realism in a computer-generated scene. one of the most popular shadow generation algorithms is shadow mapping, due to its simplicity, generality and high speed. with shadow mapping, the scene is first rendered from the view of the light, storing the depth values in a separate buffer. when the scene is then rendered from the normal viewing position, each pixel is transformed again into the light view, and its depth value compared to the depth value stored in the shadow map. if the depth value of the shadow map is nearer to the light source, the pixel is in shadow.

like all image-space algorithms, shadow mapping sufers from aliasing artifacts due to quantization and perspective projection. recently, several approaches have tried to reduce those aliasing artifacts. one particularly promising idea, perspective shadow maps, is based on perspective reparameterization of the shadow map. the original perspective shadow map technique suffers from a few drawbacks, which result mostly from the fact that chosen perspective mapping is based on the observer projection:

1. the lights have to be transformed into post-perspective space, and frequently change their type (from point to directional and vice versa, or from a normal to an “inverted” lightsouce). thinking in this particular post-perspective space is not intuitive.
2. the mapping to post-perspective space has a singularity, which causes problems with shadow casters on or opposite the singularity. the practical solution to this problem is to “move the” viewpoint backwards for the shadow map generation until all relevant objects are included in fron of the singularity. however, this reduces the shadow-map quality.
3. due to a number of special cases, the implementation is quite involved.
4. the increase in shadwom map resolution near the viewer comes at the expense of a drastic reduction of resolution for distant objects. 提高近处的分辨率，远处的分辨率就牺牲了。

in this paper, we introduce light space perspective shadow maps (LiPSM), a new shadow mapping technique based on a variable perspective mapping specified in light space (see figure 1 for an exmaple).

Figure 1: Light space perspective shadow maps (LiSPSM) (left) and the corresponding warped light view (including the eye frustum) (right). Note the high shadow detail both for near and distant objects.

the advantage of the technique is that the chosen perspective mapping has no relevant singularites, allows treating all lights as directional lights and does not change the direction of the light sources. therefore, most of the problems of perspective shadow maps are avoided. one of the most important contributions of this paper is a thorough error analysis of all shadow mapping techniques based on perspective reparameterizations. this analysis is then expolited to derive optimal parameters for our LiPSM technique.

an important insight is that perspective shadow maps trade near resolution against large shadow errors in distant objects. in contrast, our method can be tuned to distribute the error equally or in a user-specified way among different distance regions. light space perspective shadow maps are roubust, as fast as standard shadow maps and simple to implement. their major limitation is inherent to all methods based on reparameterizing the shadow plane, i.e., they can only deal with perspetive aliasing. dealing with more kinds of shadow aliasing usually requires a potential costly scene analysis per frame.

The remainder of the paper is organized as follows: after
a discussion of previous work in Section 2, we describe
the LiSPSM technique in Section 3. In Section 4, we provide
a thorough error analysis of uniform, perspective and
light space perspective shadow maps and show how to optimally
calculate the free parameter of LiSPSMs (n), which
controls the distribution of aliasing error. Results are given
in Section 5, and conclusions and ideas for future work in
Section 6.

2. previous work
   the two most important categories of shadow algorithms are shadow volumes [Cro77] and shadow mapping[Wi178]. shadow volumes work in object space and are therefore potentially more accurate, but place a high burden on geometric and pixel processing and require well tesselated objects. therefore, interest in shadow mapping techniques is large, since they work in image space and require only one additonal rednering pass per light.

most of the publications dealing with shadow maps try to solve the associated aliasing artifacts. percentage closer filtering [RSC97] alleviates reprojection problems by filtering. a number of papers have tried to solve perspective aliasing due to the perspective view-frustum projection. the most prominent of these is the perspective shadow map method by Stamminger and Drettackis[SD02], which tries to remove perspective aliasing by subjecting the shadow map to the same perspective transform as the viewer. despite its drawbacks, this paper has inspired and opened the door to more general shadow map reparameterization approaches, of which we present one in this paper. another way to reparameterize the shadow map is to tilt or warp the shadow plane directly. Recent approaches propose to combine shadow maps with shadow volumes or other primitives [SCH03, GLY∗03]. These techniques can potentially be combined with the light space perspective shadow maps
presented in this paper.