Fast And Reliable Space Leaping For Interactive Volume Rendering by Ming Wan, Aamir Sadiq, Arie Kauf

1
Fast And Reliable Space Leaping For Interactive
Volume RenderingbyMing Wan, Aamir Sadiq, Arie
Kaufman

• Presented by
• Allan Spale, CAVERN Viz Workshop, May 2004

2
Overview

• Algorithm that accelerates ray casting during
  interactive navigation in a complex volumetric
  scene
• Pixel depths derived with a fast cell-based
  reprojection algorithm during navigation
• Remaining pixel depths determined by precomputing
  the distance from each empty voxel to its nearest
  object boundary
• Distance-From-Boundary (DFB) jumping
• Provide suitable solution to new incoming objects
  problem during navigation
• Updates depth pixels where new objects are exposed

3
Background and Definitions

• Volumetric Data 3D grids of voxels
• Cell 3D volumetric region bounded by 8
  neighboring grid vertices (voxels)
• Algorithm works with cubic, regular, and
  rectilinear grids
• Utilizes 2 cell buffers and frame buffers
  (previous and current frame)
• Cell buffer Stores visible object cell
• Depth buffer Stores depth of each pixel,
  generated by the previous cell buffer
• Hole pixel Pixel not covered by any reprojected
  point
• Accelerated ray casting algorithm
• Use cell-based reprojection to get pixel depths
  from previous frame
• Perform DFB jumping at a hole pixel to detect its
  depth
• Detect correct depth values from step 1 with
  incoming objects
• Skip over empty regions along each ray

4
Cell-Based ReprojectionReproject the Cell
Center (Fig 1a)

• E viewpoint, f distance from E to image plane
• Given cell C and cells center point V
• Calculate projected position of P of V on image
  plane as
• The intersection between image plane
• Line L passing through V and E

5
Cell-Based ReprojectionDetermine Reprojected
Region (Fig 1a)

• Line L is perpendicular to the image plane
• Bounding sphere of cell C is a circle centered at
  P with radius S
• Calculate radius of S
• There is tangent plane T of cell C passing
  through viewpoint E
• A is a tangent point
• B is projection of point A on the image plane
• Distance of s between B and P
• Distance of r between A and V
• With right triangles, can use equation
• s r f / (d2 r2 ) 1/2

6
Cell-Based ReprojectionUpdate the Depth Buffer

• Determine pixels that the reprojected cell may
  occupy
• If occupied pixels depth value is available in
  the buffer
• Compare it with the new distance l from the cell
  being reprojected
• If depth value is greater than l
• The value l will replace old depth value
• Advantages
• Safe estimation of jumping distance without
  losing pixels
• Moderate impact on reprojection cost
• Less expensive than splatting because each cell
  reprojection covers a larger area on the image
  plane

7
Accelerated Ray Casting

• If a pixel has its depth value in the depth
  buffer, ray casting quickens because it can jump
  to the nearest object voxel
• Otherwise, perform DFB-jumping to skip over empty
  ray segment
• With volumes, camera is located in empty region
• Voxels DFB value is view-independent and can be
  calculated as a preprocessing step
• Whenever the ray sampling point occurs in an
  empty region, DFB jumping may be applied at that
  position
• DFB is not applicable for a camera inside a
  detailed scene since DFB would skip new objects

8
New Incoming Object Detection

• Find pixels at edge of the reprojection region in
  current image
• Recalculate depth of each peripheral pixel by ray
  casting and performing DFB until first voxel
  found
• New depth lt original depth ? mark pixel to
  indicate new object detection
• Check each marked pixels 8 neighboring pixels

• Look for pixels with depth values in the depth
  buffer but have not been accessed yet
• Recalculate their depth values
• Update depth buffer with new value
• Mark the pixel so that it indicates the detection
  of a new object
• Go to step 3 until no more pixels can be marked

9
Results

• Implementation
• 16 processor SGI Power Challenge
• Task assigned cell buffer entries and pixels to
  each processor in interleaved scanline order for
  reprojection and DFB-jumping
• Detection of new object was not parallelized
  because execution time was short
• Datasets
• Phantom pipe 512 x 512 x 107
• Colon (from CT data) 512 x 512 x 411

10
Summary

• Highlights
• Fast, reliable space leaping method to accelerate
  ray casting for large volumes
• Combines temporal and object space coherence
• Pros
• Notable speedup in rendering
• Usable for all volume grid types
• Cons
• Generic algorithm but works well in empty scenes
• Questionable image quality (figure 5) or bad PDF
  image
• New object detection tends to fail when view
  changes too much between adjacent frames