Real-time Graphics for VR

1
Real-time Graphics for VR

• Chapter 23

2
What is it about?

• In this part of the course we will look at how to
  render images given the constrains of VR
• we want realistic models,
• eg scanned humans, radiosity solution of the
  environment etc (lots of polygons/textures)
• we need real-time rendering
• over 25 frames per second
• often maintaining the frame rate is more
  important than image quality

3
How can we accelerate the rendering?

• Using graphics hardware that can do the intensive
  operations in special chips
• as processing power increases so do user
  expectations
• Fine tuning the models
• removing overlapping parts of polygons
• removing un-needed polygons (undersides etc)
• replacing detail with textures
• Improving the graphics pipeline This is what we
  will concentrate

4
Making the most of the graphics hardware

• Know the strengths and limitation of your
  hardware
• multipass texturing
• display lists, etc
• Dont compromise the portability, if software to
  be used on other platforms
• Be aware of the rapid changes in technology
• eg bandwidth vs rendering speed

5
Whats wrong with the standard graphics pipeline

• It processes every polygon therefore it does not
  scale
• According to the statistics, the size of the
  average 3D model grows more than the processing
  power

6
We can use several acceleration techniques which
can be broadly put into 3 categories

• Visibility culling
• avoid processing anything that will not be
  visible in (and thus not contribute to) the final
  image
• Levels of detail
• generate several representations for complex
  objects are use the simplest that will give
  adequate visual result from a given viewpoint
• Image based rendering
• replace complex geometry with a texture

7
Constant frame rate

• The techniques above are not enough to assure it
• We need a system load management
• it will try to achieve an image with the best
  quality possible given within the give frame time
• if there is too much load on the system it will
  resolve to drastic actions (eg drop objects)
• its an NP complete problem

8
The Visibility Problem

• Select the (exact?) set of polygons from the
  model which are visible from a given viewpoint
• Average number of polygons, visible from a
  viewpoint, is much smaller than the model size

9
Visibility Culling

• Avoid rendering polygons or objects not
  contributing to the final image
• We have three different cases of non-visible
  objects
• those outside the view volume (view volume
  culling)
• those which are facing away from the user (back
  face culling)
• those occluded behind other visible objects
  (occlusion culling)

10
Visibility Culling
11
Visibility methods

• Exact methods
• Compute all the polygons which are at least
  partially visible but only those
• Approximate methods
• Compute most of the visible polygons and possibly
  some of the hidden ones
• Conservative methods
• Compute all visible polygons plus maybe some
  hidden ones

12
View volume culling

• Assuming the scene is stored into some sort of
  spatial subdivision
• We already saw many earlier in the course, some
  examples
• hierarchical bounding volumes / spheres
• octrees / k-d trees / BSP trees
• regular grid

13
View volume culling

• Compare the scene hierarchically against the view
  volume
• When a region is found to be outside the view
  volume then all objects inside it can be safely
  discarded
• If a region is fully inside then render without
  clipping
• What is the difference with clipping?

14
View volume culling against a bounding volume
hierarchy
15
View volume culling against a space partitioning
hierarchy
16
View volume culling

• Easy to implement
• A very fast computation
• Very effective result
• Therefore it is included in almost all current
  rendering systems

17
Back-face culling

• Simplest version is to do it per polygon
• just test the normal of each polygon against the
  direction of view (eg dot product)
• More efficient methods operate on clusters of
  polygons
• group polygons using the direction of their
  normals, make a table
• compare the view direction against the entries in
  this table

18
Occlusion culling

• By far the most complex (and interesting) of the
  three, both in terms of algorithmic complexity
  and in terms of implementation
• This is because it depends on the inter-relation
  of the objects
• Many different algorithms have been proposed,
  each one is better for different types of models
• Whats the difference with HRS