Page 11 - Developer

we have a new view of the same same y value as our desired view
LISTING 1 The core of the fragment shader (using GLSL).
scene we wish to render. We plane location, so we’ll just search a
could just use the same algorithm range of candidate screen positions
UNIFORM SAMPLER2D COLOR_TEXTURE; by progressively casting line P* along a line of constant y. In
UNIFORM SAMPLER2D DEPTH_TEXTURE; segments per pixel to generate the this case we start with our pixel’s
UNIFORM VEC4 STEREO_PARAMS; new view. However, because of the position from the right eye viewpoint
VARYING VEC2 UV; nature of stereo reprojection, we (in green) and iterate in the -x
VOID MAIN() can simplify things greatly. Figure direction (to yellow and orange). So,
{ 4 shows the situation for a desired rather than casting a ray and then
VEC2 OUT_TEX_COORDS = UV - VEC2(STEREO_PARAMS.X, 0.0); location in the right eye view (in computing a view plane position
VEC2 REPROJ_TEX_COORDS = UV; green). We create sample points to look up a depth value, we’ll just
FLOAT DIR = SIGN(STEREO_PARAMS.Y); along a ray from the right viewpoint use the depth at the current P* to
FLOAT ORIG_X = UV.X; through the view plane (in this reproject into the right eye view. If
FLOAT SHIFT = STEREO_PARAMS.Z; case, the convergence plane) its position is close to our desired
FLOAT STEP = 0.0625*STEREO_PARAMS.W; location for our current pixel. Since location then we’ll stop. We’ll use
INT INDEX = 0; the view separation is only in the that depth as our final depth to look

WHILE (INDEX <= 16) x-direction, we end up only casting up the color value for our pixel.
{ the ray along x and z. The y-value of This raises the question of
REPROJ_TEX_COORDS.X = ORIG_X + SHIFT; the ray will always be 0. how we perform our reprojection—
FLOAT L = GETLINEARDEPTH(DEPTH_TEXTURE, REPROJ_TEX_ For the sake of clarity, we start namely, how do we use a view
COORDS); our sampling in Figure 4 at the plane position P_left in the left
FLOAT DEPTH_ADJUST = STEREO_PARAMS.X + STEREO_PARAMS.Y convergence plane, though we view and the depth to determine
/ L; usually will want to start between the corresponding view plane
FLOAT TEST = SHIFT + DEPTH_ADJUST; the eye point and the convergence position P_right in the right view?
IF (DIR*TEST >= 0.0) plane to get some negative We can simplify the problem down
{ parallax. For each sample point (in to one of finding the displacement

OUT_TEX_COORDS.X = ORIG_X - DEPTH_ADJUST; yellow), we cast a new ray from the along the convergence plane,
BREAK; left eye, determine the view plane between the left view and the right
} position for that ray, and look up view (see Figure 6). This is a case
SHIFT = SHIFT + STEP; INDEX = INDEX + 1; the depth by mapping that view of computing a ratio of similar
} plane position to texture space. If triangles: The distance between
GL_FRAGCOLOR = TEXTURE2D(COLOR_TEXTURE, OUT_TEX_COORDS); this depth is less than the depth P_left and P_right is to the distance
}
of the sample point, we stop and between the two view positions as
use an interpolation of the sample the distance from the convergence
information in Brawley and the depth map, you use that point’s depth and left eye depth plane is to the distance from the
Tatarchuk’s original ShaderX article penetration point to determine the to get our final depth value, which view positions. This gives us the

[see Reference 4]. However, the color seen from that view direction. we use to reproject from the right final equations in Figure 6.

general idea is that when you Stereo reprojection, at its viewpoint to get our final color. Since this is using a ratio of
render the surface you wish to heart, is essentially the same However, because we are only distances, this will work just as
apply parallax mapping to, for each process. We have a color image searching in the xz directions, well whether we’re in NDC space
pixel you cast a line segment along and its corresponding depth we can think of this in a different or texture space, but it assumes
the view direction at progressive information (albeit using a way—see Figure 5. In this case, we that the depth we have is linear.
depths into the depth map. When perspective projection rather know the corresponding view plane However, if we use standard
the line segment has penetrated than an orthographic one), and position in the left view will have the perspective projection, the depth

FIGURE 4: Stereo FIGURE 5: Stereo
reprojection using reprojection using
ray-casting. pixel match.

WWW.GDMAG.COM 9