Slides PPT 4,747KB

Download Report

Transcript Slides PPT 4,747KB

The-Kiet Lu
Kok-Lim Low
Jianmin Zheng
Graphics Interface 2009
1

Rendering high-detail surface (>100M points) is
inefficient when vertices are independently processed
one-by-one in real-time
• Redundant processing
 Occlusion
 LOD
 # of pixels vs. # of vertices
 1024x1024 pixels vs. 368 millions points
 Memory size
 1 vertex requires > 6 bytes (un-compressed)
 St. Mathew (368 M) > 2Gbs

Simplification – slow
• Requires surface reconstruction first
2
 Hybrid
approach
• Object-space and screen-space solution
• Display 3D models by per-pixel ray-casting a set
of 2D height fields (displacement mapping)
3
4
 1)
2D Height Field Construction
• Decompose points using octree
• Generate height field maps
• Generate height field bounding boxes
 2)
Height Field Rendering
• Rasterize bounding box.
• For each fragment rasterized:
 Transform viewing ray to height field map coordinate
space
 Ray-casting — compute ray-surface intersection.
5
 Decompose
points using an octree so that each
cell contains points forming a valid height field
6
 Using
PCA to determine Height Field domain
plane. Height Field map is generated as
displaced distance between the Surface and
the PCA plane.
7
 The
faces of the octree cells
that bounds the height filed
surface are added into
Vertex buffer with
appropriately assigned
texture coordinates.
8
 Each
fragment, the viewing ray is transformed
from world coordinate frame into locals of the
corresponding 2D height map.
9

We compute intersection by marching along the
viewing ray from initial location
v
to the closet point above the
Height Field map.
Xi = Xi-1 + v x ∆h

Culling Test:
A simple Texture coordinate check
to ensure the ray indeed intersects
the current height map – not the
neigh borings.

10
 Hybrid
approach
• Optimize
 Automatic Screen-space occlusion culling
 Back Face culling
 Early Z-culling
 Memory consumption
 6 bytes vertex vs. 8-bit depth buffer
 Texture compression
• Limitation
 Performance subjects to screen resolution
11
A main advantage of
our method is
its simplicity in
dealing
with silhouettes
correctly.
12
1 iterations
4 iterations
8 iterations
13
14
15
16
17

Comparison: point-based Mar2007 : 60M per second –
our method results with 368M (St. Mathew) with 12 Fps
on GeForce Go 7900 GPU - Intel Core 2 Duo 2.4 GHz..
18


We have presented a new approach for fast
visualization of highly detail 3D models by
decomposition 3D model into Height Fields and
render these Height Fields using image-space raycasting algorithm that offers significant
performance improvement compared to previous
point-based approach.
Our current limitation is that it is not efficient for
flat and smooth surfaces as well as not applicable
for dynamic or deformable surfaces – our current
research topic.
19