S3 Graphics. ... depth accuracy in the camera view and thus reduces two categories ... to draw the scene from the usual camera viewpoin by decoding the.
Reconstructable Geometry Shadow Maps Qinghua Dai S3 Graphics.
Baoguang Yang S3 Graphics.
Jieqing Feng Zhejiang University
Abstract Traditional shadow maps[Williams 1978] suffer from two main issues, aliasing errors and depth bias problems. We present a novel algorithm, the reconstructable geometry shadow maps (RGSM), as a solution to both problems. It is capable of producing sub-pixel depth accuracy in the camera view and thus reduces two categories of aliasing errors, the perspective aliasing errors and the projective aliasing errors. This goal is achieved by using the depth values faithfully reconstructed from the corresponding occluding triangles instead of taking point samples from the depth map. The experimental result also shows RGSM requires a much smaller depth bias value than other shadow mapping techniques.
RGSM Overview The RGSM has a similar framework as that of standard shadow maps. The first step renders the scene from the light’s point of view. The vertices information of each occluding triangles in the light unit cube are saved to the geometry shadow maps. The second step is to draw the scene from the usual camera viewpoin by decoding the proper information from geometry shadow maps, which helps to find the occluding triangle information, reconstruct the depth value of the occluding point and finally make a correct shadow determination. If there are multiple lights, separate geometry shadow maps must be used for each light. In a simple example, a pixel P in eye space is transformed into the light unit volume to test visibility from light. By sampling with its light space normalized coordinates(p.e x, p.e y ) in geometry shadow maps, the pixel can get the corresponding occluding triangle with normalized vertex coordinates v0 .g xyz,v1 .g xyz and v2 .g xyz. Thus we can reconstruct the corresponding occluding point Pd within the triangle by the following equation. `
x Pd .e z = p.e
p.e y
0 x ´ v0 .e 1 @v1 .e x v2 .e x
v0 .e y v1 .e y v2 .e y
1−1 0 1 v0 .e z 1 zA 1A @v1 .e 1 v2 .e z
(1)
By comparing Pd .e z with the depth value of pixel P, the shadow determination of pixel P can finally be made. For more complex cases, a consistency test is used to find the appropriate occluding triangle which is the closest to the light source. To solve pixel-texel mismatch problem, we use a combination strategy to find all possible occluding triangles. We assume all closestto-light triangles are rasterized on geometry shadow maps. If this requirement is not satisfied, this approach may still show its own aliasing problem. Warping and subdividing [Stamminger and Drettakis 2002 etc.] shadow techniques can be used with RGSM as long as the linearity of geometry surface can be guaranteed.
Implementation and Results We implemented RGSM on a Geforce 8800GT using DirectX10 with two versions, an RGSM(basic) and an RGSM(adjacency). The former makes use of geometry information of the central triangle and the latter uses geometry information from the central and adjacent triangles. We compare the shadow image quality of the Copyright is held by the author / owner. I3D 2008, Redwood City, California, February 15–17, 2008. ACM 978-1-59593-983-8/08/0002
Figure 1: Both anti-aliasing(1st row) and reducing depth bias(2nd row) effects of reconstructed geometry shadow maps. First row from left to right: traditional shadow maps (1K × 1K), RGSM (1K × 1K) and shadow volumes. Second Row from left to right: traditional shadow maps (bias=1e-3), standard shadow maps (bias=1e-6) and the RGSM (bias=1e-6) with the same resolution 512 × 512.
RGSM with the traditional shadow maps using a reference image of shadow volumes. By calculating the differences around the shadow border areas, the result shows the RGSM(adjacency) can achieve shadow volume like quality (> 90%) when the texture resolution is still relatively low (< 1.5K × 1.5K). The RGSM(basic) has a faster speed but achieves the same shadow quality as RGSM(adjacency) with a higher texture resolution cost. At the premise of the same shadow quality level, the memory consumption of RGSM(adjacency) is 56% and RGSM(basic) is 16% of that of traditional shadow maps. Both RGSMs have much higher speed than shadow volumes, but slower speed than traditional shadow maps. In our depth bias test, the RGSM requires only a small bias value less than 3e-6 which enables rendering image without false self-shadowing and false non-shadowing artifacts. Meanwhile, the bias value keeps almost constant across different object scenes, even with light source or camera movement.
Conclusion and Future work A new shadow rendering technique is presented for reducing aliasing and improving depth bias issues in traditional shadow map techniques. The future work include weakening the condition of consistency test and making more improvements in term of memory consumption.
References R EEVES , W., S ALESIN , D., AND C OOK , R. 1987. Rendering antialiased shadows with depth maps. Computer Graphics (In Proceedings of ACM SIGGRAPH 87) 21, 283–291. S TAMMINGER , M., AND D RETTAKIS , G. 2002. Perspective shadow maps. In Proceedings of ACM SIGGRAPH 2002, ACM Press/ ACM SIGGRAPH, J. Hughes, Ed.