Interactive Methods in

Scientific Visualization

Terrain Rendering
Christian Dick
[email protected]

Computer Graphics & Visualization Group, Technische Universität München, Germany

IEEE Pacific Visualization Symposium 2009

Beijing, China
 Outline
• Terrain Data and typical Storage Requirements
• Application Areas and their Requirements
• Overview of Continuous Level of Detail
Terrain Rendering Techniques
• Texture and Geometry Compression
• Overlaying 2D Vector Data
• Demo

Updated slides will be availabe at http://wwwcg.in.tum.de/Tutorials/PacificVis09

 Terrain Data
Orthophoto Height Field (DEM)

Geo Data © Landesamt für Vermessung und Geoinformation Bayern

 Triangle Mesh

Geo Data © Landesamt für Vermessung und Geoinformation Bayern

 Textured Triangle Mesh

Geo Data © Landesamt für Vermessung und Geoinformation Bayern

 State of Utah, USA
Texture: 1m
Height Field: 5m
Extent: 460km x 600km
Raw Data: 790 GB
(China or USA: 27 TB)
 Vorarlberg, Austria
Texture: 12.5cm
Height Field: 1m
Extent: 56km x 85km
Raw Data: 860 GB
(China or USA: 1690 TB)
Geo Data © Land Vorarlberg
Geo Data © Land Vorarlberg
Geo Data © Land Vorarlberg
Terrain Rendering – Application Areas
•Games
Games
– High, constant frame rates
– Quality advantageous but not prioritized
Simulators
•Simulators
– Constant frame rates (~30 fps)
– High degree of realism and detail, optionally stereo
Geographic
•Geographic Information
Information Systems
Systems (GIS)
(GIS)
– Interactive frame rates advantageous (15+ fps)
– High resolution and precision
 Terrain Rendering Principles
• How to render TB-sized datasets on standard
PC hardware?
– Limited memory bandwidths and capacities
– Limited rendering throughput (~350 M/s)
– Brute force not possible …
 Terrain Rendering Principles
• Only a small amount of data is visible per view
Perspective Projection
Level of Detail

Limited Field of View

View Frustum Culling
 Terrain Rendering Principles
• Out-of-core data streaming
Memory Hierarchy
Several TBs 8 GB 1 GB
~ 100 MB/s ~ 2 GB/s
Main Graphics
HDs
Memory Memory

Entire Data Set Data within a certain Data within the

Prefetching Region View Frustum

Hide Disk Access Latencies

Data Prefetching

Resource Allocation/Deallocation Reduce Memory Capacity

Caching Strategies and Bandwidth Requirements
CPU & GPU Memory Management Data Compression
 CLOD Terrain Rendering
•Dynamic
Dynamicremeshing
remeshing
– ROAMing Terrain: Real-time Optimally Adapting Meshes
[Duchaineau et al., 1997]

•Tile-based multiresolution
Tile-based mesh
multiresolution representation
mesh representation
– Rendering Massive Terrains using Chunked Level of Detail Control
[Ulrich, 2002]

Nested
•Nested regular
regular grids
grids
– Geometry Clipmaps: Terrain Rendering Using Nested Regular Grids
[Losasso and Hoppe, 2004]

•Ray-casting ofof
Ray-casting the height
the field
height field (no mesh)
– GPU Ray-Casting for Scalable Terrain Rendering
[Dick et al., 2009]

See www.vterrain.org for a list of publications …

 Dynamic Remeshing
• View-dependent, adaptive remeshing in every
frame
• LOD computation and view frustum culling per
triangle
• Remeshing is done on the CPU, mesh has to be
transferred to the GPU in every frame
• ROAM: Real-time Optimally Adapting Meshes
 Dynamic Remeshing – ROAM
• ROAM uses a Triangle Bintree Mesh for
view-dependent, adaptive meshing
– Mesh consists of isosceles right triangles
– Starting from a single triangle, triangles are
successively split from the apex to the midpoint of
the hypotenuse (longest edge bisection)
 Dynamic Remeshing – ROAM
Level 0 Level 1

Level 2 Level 3

Level 4 Level 5
 Dynamic Remeshing – ROAM
• Remeshing is done by exploiting frame-to-
frame coherence
– Mesh refinement: Split triangles using
diamond splits to avoid T-vertices
– Mesh coarsening: Merge diamonds T-vertex

Split

Merge
 Dynamic Remeshing – ROAM
• Diamond splitting rule leads to forced splits
 Dynamic Remeshing – ROAM
• Remeshing is driven by two priority queues
– Priority = Screen space error
– Split Queue: Force-split triangles with highest
priority
– Merge Queue: Merge diamonds with lowest
priority
– Uses view-independent, precomputed world-space
error bounds for the triangles in the bintree
– View-dependent priorities are obtained from these
error bounds by projection into screen-space
(Priorities have to be updated when view position changes)
 Avoiding Popping Artifacts
• Geomorphs (vertex morphing)
– Linear interpolation of the height of the center vertex
z vC
z(vC)
LC RC z  t   1  t  zT vC   t z vC  ,
zT(vC) t  0, 1
T

• Subpixel geometric screen-space error

– Use screen space error tolerance < 1 pixel
z vC
z(vC) RC
LC
< 1 pixel
zT(vC)
T
 Dynamic Remeshing
• Advantages:
– Continuous triangulation (no T-vertices, no cracks)
– Represents terrain using a minimum number of
triangles
• Disadvantages:
– High CPU load
– High load on the CPU-GPU bus
 Tile-based Terrain Rendering
• Terrain is represented at multiple levels of
details
• Each level is divided into tiles, which are
organized as a quadtree
• Each tile is meshed separately in an offline
preprocess (with respect to a prescribed
world-space error tolerance per LOD)
 Tile-based Terrain Rendering
• In each frame, the set of tiles is determined
which represents the terrain at a prescribed
screen-space error, using LOD computation
and view frustum culling (per tile)
• Only new tiles are uploaded to the GPU
(exploiting frame-to-frame coherence)
• Chunked LOD: Rendering Massive Terrains
using Chunked Level of Detail Control
[Ulrich, 2002]
– Demo with source code available at
http://tulrich.com/geekstuff/chunklod.html
 Tile-based Terrain Rendering
• Multi-resolution, tile-based terrain representation
Level 3 1x1 Tile 5122 Samples

Level 2 2x2 Tiles 10242 Samples

Level 1 4x4 Tiles 20482 Samples

Level 0 8x8 Tiles 40962 Samples

Geo Data © Landesamt für Vermessung und Geoinformation Bayern

 Tile-based Terrain Rendering
• Facts about the Tile Quadtree
Level 3
1x1 Tile – 5122 samples per tile
Level 2
2x2 Tiles
– World-space sample
Level 1 spacing  and world-
4x4 Tiles
space tile extent 512 
are doubled with each
level  (bottom-up)
Level 0
8x8 Tiles – The tiles are meshed
with a world-space
error tolerance of 
Geo Data © Landesamt für Vermessung und Geoinformation Bayern
 Tile-based Terrain Rendering
• View frustum culling and LOD computation per tile
 Tile-based Terrain Rendering
• LOD computation is based on the theorem on
intersecting lines  P   '

Pz zNear
2  
  P    tan     Pz
h  
2
  P : World-space error tolerance at point P
h: View port height in pixels
: fovy
: Screen-space error tolerance in pixels
Pz : Camera-space depth of point P

Use a tile at level  for rendering, iff

     P '     1
where P ' denotes the point of the tile's bbox
with the least depth to the camera
 Tile Selection Algorithm
• To determine the set of tiles
to be rendered, traverse the
tile quadtree in preorder
– If a tile is culled, skip all
descendants
– If a tile is not culled and its LOD
is sufficient with respect to the
prescribed screen-space error
tolerance, add this tile to the
set and skip all descendants
 Fixing Cracks between Tiles
• Tiles are meshed independently, thus cracks
(due to T-vertices and quantization errors) can
occur at the tile boundaries
• Render skirts around each tile to hide cracks

• Alternatives: Flanges, Zero-Area-Triangles

(see [Ulrich, 2002])
 Texturing
• Anisotropic texture filtering is mandatory for
terrain rendering

Mipmapping Anisotropic Filtering

Geo Data © Landesamt für Vermessung und Geoinformation Bayern

 Tile-based Terrain Rendering
• Advantages:
– Low CPU load (computations are done per tile)
– Moderate load on the CPU-GPU bus due to frame-
to-frame coherence
• Disadvantages:
– Cracks between tiles
– Higher number of triangles than necessary
 Nested Regular Grids
• Geometry Clipmaps: Terrain Rendering Using
Nested Regular Grids
[Losasso and Hoppe, 2004]
– Demo available at
http://research.microsoft.com/en-us/um/people/hoppe/
• Use a set of nested regular grids centered
about the viewer
Nested Regular Grids
• Height values are
fetched on-the-fly
from a height field
pyramid (the clipmap)
• LOD is controlled by
the spatial extent of
each grid
• View frustum culling is
realized by dividing
each grid into blocks
 Nested Regular Grids
• Update of the clipmap based on toroidal access
Update region

Viewer Position
 Nested Regular Grids
• Advantages:
– Height field can be compressed using image
compression methods
– Simple memory management
• Disadvantages:
– No exact screen space error control
– Extremely high number of triangles
 GPU-based Terrain Ray-Casting
• Ray-casting of the terrain height field
• GPU Ray-Casting for Scalable Terrain
Rendering [Dick et al., 2009]
BoxExit
BoxEntry

View Plane

Eye
Dir = BoxExit - Eye
 GPU-based Terrain Ray-Casting
• Advantages:
– Performance fully independent of the complexity
of terrain
– Higher performance and lower GPU memory
consumption than triangle-based rendering for
high-resolution height fields
• Disadvantages:
– For coarse-resolution height fields, triangle-based
rendering is faster and requires less GPU memory
 Data Compression
• Benefits:
– Reduces memory capacity requirements
– Reduces bandwidth requirements
• Favor schemes that can be decoded on the GPU
– Reduces CPU load
– Reduces CPU-GPU traffic
• Encoding generally not time-critical
– Performed in a (time-consuming) preprocess
 Texture Compression – S3TC
• S3 Texture Compression, here: DXT1, no alpha
[US Patent 6658146]
– Asymmetric, lossy block truncation code
– Standard compression scheme (DirectX, OpenGL)
– GPU renders directly from compressed data
– Divides textures into 4x4 blocks
– Assigns a fixed rate of 64 bits per block (4 bpp)
– Compression ratio 6:1 (R8G8B8)
– Use the Squish library by Simon Brown for
compression; Available with source code at
http://code.google.com/p/libsquish/
 Geometry Compression
• Compression scheme for bintree meshes
supporting GPU-based decoding
– Efficient Geometry Compression for GPU-based
Decoding in Realtime Terrain Rendering
[Dick et al., 2009]
– Underlying 2D Mesh: Lossless compression based
on a generalized triangle strip representation
– Height values: Lossy compression based on
uniform quantization
– Compression rate 8-9
(wrt triangle list representation, 32 bits per vertex)
 Geometry Compression
• Generalized Triangle Strip
– Store only one vertex per triangle
1 3 5 1 3 4 5
7

0 2
4 0 2 6
Regular Triangle Strip Generalized Triangle Strip
(0-1-2, 2-1-3, 2-3-4, 4-3-5, …)

• Construct a directed path that

– Enters each triangle exactly once
– Leaves and enters triangles across edges
– Hamiltonian path of dual graph
 Geometry Compression

5
0 3
Start

1 2 4
 Geometry Compression
• Classify triangles by
– Type of the entering/leaving edge (A, B, C)
– Winding of the path (L, R)
Type A: Type B: Type C:
Cathetus to cathetus Cathetus to hypotenuse Hypotenuse to cathetus
 Geometry Compression
• Construct path during diamond splitting
– Initial Mesh
 Geometry Compression
• Construct path during diamond splitting
– Replacement System
Geometry Compression
Geometry Compression – Encoding
• For each triangle
– Store type (A,B,C) New vertex

– Winding (L,R) can be inferred

– Store height value of new vertex
• Bitrate
– 2 bits for triangle type
Already known vertices
– Variable (per tile) #bits
for height value
Geometry Compression – Decoding
• Encoded mesh:
CL, CR, AL, AL, …, BR, BL, AL + height values

• Can be decoded directly on the GPU

 Overlaying 2D Vector Data
• 2D Vector Data
– Polyline and polygonal vector data
– Roads, trails, villages, land use, …

• Overlaying onto the 3D terrain

– For the visualization, the 2D vector data have to be
mapped onto the 3D terrain surface …
2D Vector Data
3D Terrain
Vector data overlaid onto the terrain
 Overlaying 2D Vector Data
•Geometry-based
Geometry-based
– Render vector data as 3D geometry (lines, triangles)
– Problems: Z-fighting, terrain LOD adaptation
•Texture-based
Texture-based
– Rasterize vector data into texture, overlay texture
– Problems: Resolution, GPU memory consumption
ShadowVolume”
•“Shadow volume approach
Approach
– Efficient and Accurate Rendering of Vector Data on
Virtual Landscapes [Schneider and Klein, 2007]
 Thanks for your attention!
Online Demo: Tile-based Terrain Rendering

• Data Set: State of Utah, USA

– Texture / Geometry Resolution: 1m / 5m
– Spatial Extent: 460km x 600km
– Raw Data Volume: 790GB (Compressed 175GB)

• System: Notebook equipped with

– Intel Mobile Core 2 Duo T7500, 2.2GHz
– NVIDIA GeForce 8600M GS, 256MB video memory
– 2GB of RAM
– External Harddisk, connected via USB
Updated slides will be availabe at http://wwwcg.in.tum.de/Tutorials/PacificVis09