1 eg2000 recent advances in visualization of volumetric data ken brodlie jason wood university of...

83
1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

Post on 22-Dec-2015

217 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

1EG2000

Recent Advances in Visualization of Volumetric

Data

Recent Advances in Visualization of Volumetric

Data

Ken BrodlieJason Wood

University of Leeds

Page 2: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

2EG2000

ApplicationsApplications

Simulations from computational science

Medical imaging

Page 3: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

3EG2000

ApplicationsApplications

Voxel-Man Anatomical Atlas

University of Hamburg

Page 4: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

4EG2000

Scalar Data in 3DScalar Data in 3D

Rectilinear

Curvilinear

Unstructured

Page 5: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

5EG2000

Reference ModelReference Model

Haber-McNabb model describes visualization as a pipeline:

Interpolatethe samples

Apply a technique

Draw thegeometry

Data Enrich Map Render

Page 6: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

6EG2000

Structure of the TalkStructure of the Talk

Data Enrich Map Render

Page 7: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

7EG2000

Data Enrichment : Interpolation

Data Enrichment : Interpolation

Rectilinear data:– piecewise trilinear– 7 linear interpolationsF(x,y,z) = a + bx + cy +dz

+ eyz + fzx + gxy + hxyz

– piecewise tricubicF(x,y,z) = 64 terms!

.. recently used by Cohen-Or et al (2000) for smoothing boundary surfaces

f101

f000

f001

f100

f111

f011

f110f010

..but deceptively complex,eg surface F = constantis conic surface

Page 8: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

8EG2000

Data Enrichment: Interpolation

Data Enrichment: Interpolation

Medical imaging: cubic interpolation in z-direction only could make sense

Unstructured data:– piecewise linear within

each tetrahedronF(x,y,z) = a + bx + cy +

dz– other approaches -

radial basis functions, Shepard methods, … (see Nielson review papers)

.. surface F = constantreally is simple

Page 9: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

9EG2000

MappingMapping

Three approaches:– isosurfaceF(x,y,z) = k

section through dependent variable

– sliceF(x,y,z) such that (x,y,z)

Psection through

independent variable

– volume renderingF(x,y,z) mapped to

volume of translucent gel

Page 10: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

10EG2000

RenderingRendering

Geometry rendered using computer graphics techniques

– polygon rendering and ray casting

Note interaction with mapping step

– isosurface geometry approximation disguised by clever rendering ..

.. but care needed orelse..

Page 11: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

11EG2000

Structure of TalkStructure of Talk

Data Enrich Map Render

Isosurface Slice Volume Render

Page 12: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

12EG2000

Isosurfacing - Classical Approach

Isosurfacing - Classical Approach

Marching CubesLorensen and Cline (1987)– Rectilinear grid of cells– Estimate F(x,y,z) by

trilinear interpolant within each cell

– Calculate intersections of isosurface with cell edges

– Approximate with simple triangulation in interior

– 15 canonical configurations Marching Tetrahedra

– Unstructured data

Page 13: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

13EG2000

Isosurfacing - Improving Marching Cubes

Isosurfacing - Improving Marching Cubes

Surface representation:

– naïve interior representation can leave holes

– increase robustness and accuracy

– preserve shapes

Performance– few cells contribute

to surface– how can we find

them quickly?

Page 14: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

14EG2000

Structure of TalkStructure of Talk

Data Enrich Map Render

Isosurface Slice Volume Render

Robustness

Accuracy

Topology

Shape

MarchingCubes

Surfacerep’n

Perform-ance

Page 15: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

15EG2000

Surface representation: Robustness

Surface representation: Robustness

Getting rid of the dreaded holes!

Full 256 case table allows a robust algorithm

Page 16: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

16EG2000

saddle pointof bilinearinterpolant

Surface representation: Topological correctnessSurface representation: Topological correctness

Ambiguous faces– asymptotic decider of

Nielson and Hamann (1992) - resolves ambiguity by matching the bilinear face behaviour

– decided by saddle point value

– subcases for each of the 6 ambiguous cases of the original 15 -making 27 cases in all

– interior triangulation still simple

Page 17: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

17EG2000

Surface representation: Topological correctnessSurface representation: Topological correctness

Ambiguous interior behaviour

– MC case 5 – increase the data

values until surfaces touch and then a tunnel appears

– Natarajan (1994) resolved ambiguity by body saddle value

body saddle = 3D equivalent of 2D saddle, where all first derivatives zero

Page 18: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

18EG2000

Surface representation:Topological correctnessSurface representation:Topological correctness

Definitive work by Chernyaev (1995)

– 15 basic canonical configurations

– Ambiguous faces divide 6 into 18 subcases - totalling 27 cases

– Ambiguous interiors (tunnel or no tunnel) divide 6 into 12 cases - totalling 33 cases

Marching Cubes 33

Page 19: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

19EG2000

Surface representation: Accuracy

Surface representation: Accuracy

Can we represent the interior surface within a cell more accurately?

One approach is to approximate the trilinear surface with a curved surface

– Hamann et al (1997) used triangular rational-quadratic Bezier patches

Page 20: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

20EG2000

Surface representation: Accuracy

Surface representation: Accuracy

Another approach is to refine the polygonal approximation

Start from contouring– For bilinear interpolant,

contour is hyperbolic arc (PRQ)

– Usually approximated by single line (PQ)

– Suppose we use two lines - how do we best do this?

– Shoulder point (R) is point on hyperbola furthest from chord

Page 21: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

21EG2000

Surface representation: Accuracy

Surface representation: Accuracy

Lopes (1999) carried this over to isosurfacing

Greater accuracy from:

– adding shoulder points in faces

– adding inflection points in interior

inflection points = points on surface which are saddle points on planes parallel to cell faces

Page 22: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

22EG2000

Surface representation: Accuracy

Surface representation: Accuracy

Inflection points mark significant changes in topology

Also characterise the nature of tunnels

Page 23: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

23EG2000

Surface representation: Accuracy

Surface representation: Accuracy

An application where triangle count is unimportant is manufacturing

Bailey (2000) describes refining triangles by comparing value at mid-points of sides and centroid with isosurface value

Similar recent work by Cignoniet al (2000) - comparison based on distanceto isosurface

Page 24: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

24EG2000

Surface representation: Rendering

Surface representation: Rendering

Gouraud or Phong shading will improve appearance of triangular mesh

Better results by calculating gradient vector of function

– rectilinear grid: central differences at gridpoints (= normal to best fit plane)

– unstructured grid: best fit plane to data - with distance weighting

Page 25: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

25EG2000

Surface representation: rendering

Surface representation: rendering

At this conference:– Neumann et al

describe new method which fits linear function at each grid point using distance weighting

F(x,y,z) = a + bx + cy + dz

choosing a, b, c, d to best fit at each grid point

– in context of volume rendering, but could be applied to isosurface shading

Page 26: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

26EG2000

Surface representation: Direct Surface RenderingSurface representation:

Direct Surface Rendering

Jones and Chen (1995) suggested direct ray tracing of surface

– High quality images, with shadows

– No intermediate geometry, view dependent...but is this a problem?

Parker et al (1999) follow up with interactive ray tracing

– exploits parallelism on 64 processor SGI Reality Monster

Page 27: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

27EG2000

Isosurfaces - Shape-basedIsosurfaces - Shape-based

Early approach was to contour 2D slices -and stitch them together into 3D surface- but problems in handling branches

– Jones and others have suggested expressing 2D slices as distance fields (distance of point from contour line)

– Stacking the distance field slices into a volume and isosurfacing solves branching problems

1

-1

-2

1

-1

-2

-3

-1

-2

-1

-2

-3

distance field characterises theshape of the contour

Page 28: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

28EG2000

Isosurfaces - Shape-basedIsosurfaces - Shape-based

Now substantial interest in shape-based interpolation which uses the distance field idea

Udupa, University of Pennsylvania

Page 29: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

29EG2000

Isosurfaces - Shape-based Example

Isosurfaces - Shape-based Example

Visualizing Circle of Willis is a challenging problem

Isosurfacing the raw data not successful

Segmentation can identify the structure, but on closer inspection boundary surface not smooth

Page 30: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

30EG2000

Isosurfaces - Shape-basedIsosurfaces - Shape-based

Improved results from:

– extract contours– smooth– create distance field– stack distance fields– interpolate between

slices to increase resolution

– isosurface the distance field data

Page 31: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

31EG2000

Structure of TalkStructure of Talk

Data Enrich Map Render

Isosurface Slice Volume Render

Robustness

Accuracy

Topology

Shape

MarchingCubes

Surfacerep’n

Perform-ance

Structured

Unstructured

Page 32: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

32EG2000

Surface Extraction - Performance

Surface Extraction - Performance

Performance issues with the Marching Cubes algorithm.– Visits and tests every cell in the

data set for surface intersections.

– Produces large numbers triangles, some of them extremely small.

Page 33: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

33EG2000

Performance Performance

Structured data– Implicit connectivity between cells.– Data stored in simple 3D array.

Unstructured data– Connectivity explicitly described.– Data not necessarily stored in a

consecutive, particularly true of adaptive meshes.

Page 34: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

34EG2000

Structured data - OctreeStructured data - Octree

Can take advantage of implicit cell ordering when constructing search structures.

Octree - recursive sub-division of geometric space.– Trees are only optimal

when data size is 2nx2nx2n

Page 35: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

35EG2000

Structured data - Branch on need octree

Structured data - Branch on need octree

Branch On Need Octree (BONO)– Wilhelms and Van Gelder 1992– reduces tree overhead by minimising the

number of subdivisions, particularly at the lowest levels.

Page 36: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

36EG2000

Structured Data - Multi-resolution

Structured Data - Multi-resolution

Multi-Resolution Approach– Reduce the number of cells to visit by

combining cells with similar values into larger cells.

– Leaves cracks in surface where areas of different resolution join.

– Gives a mesh of varying resolution across the data set.

Page 37: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

37EG2000

Multi-resolutionMulti-resolution

Tetrahedral Framework approach - Zhou,Chen,Kaufman 1997– Check data is ((2n+1)x(2n+1)x(2n+1)), pad

with blank data if required.

– Take centre point and form 6 pyramids with the faces as the base of each.

– Recursively sub-divide each pyramid into tetrahedra until at original cell level

Page 38: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

38EG2000

Multi-resolutionMulti-resolution

Tetrahedral Framework approach - cont’d.– Use some method to allow re-

combination of low level tetrahedra using an error estimate and ensuring no hanging nodes.

– Use Marching Tetrahedra to construct iso-surface.

Many others working in this area such as: Cignoni 1997, Ohlberger 1998, and others.

Page 39: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

39EG2000

Different approach for unstructured data

Different approach for unstructured data

Structured data approaches use a geometric approach to simplification that is not easy to apply to unstructured grids.

Many of the approaches for unstructured data construct value based search structures.

Page 40: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

40EG2000

Performance - Extrema GraphsPerformance -

Extrema Graphs

Itoh & Koyamada 1995.

Find maxima and minima within data set

– reduce to single cells

Join maxima and minima points by arcs of connected cells

– Direct arc where possible, otherwise use polygonal arcs.

Page 41: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

41EG2000

Extrema GraphsExtrema Graphs

Extrema Graphs cont’d.– if direct connection fails due to

crossing boundary of data set pick different point to connect to.

If no cells on direct path then must use polygonal arc, expensive.

Page 42: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

42EG2000

Extrema GraphExtrema Graph

Extrema Graph cont’d.– to create isosurface, search through

boundary lists and connection lists looking for seed points.

– Use a surface propagation algorithm to grow surface from these seed points.

Page 43: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

43EG2000

Performance - Kd-Trees

Performance - Kd-Trees

Livnat, Shen & Johnson 1996– previous approaches create

structures ordered by either min or max, or have a structure for each.

– Combines min max lists into single tree data structure.

– Kd-Tree is multi-dimensional binary tree.

Page 44: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

44EG2000

Kd-TreesKd-Trees

Representation using Span Space– allows us to geometrically understand

range based methods.

– plot each range as a single point in span space.

– for a given threshold, T, can easily show which cells are active.

Page 45: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

45EG2000

Kd-TreesKd-Trees

Constructing the Kd-tree– First find min max range of each

cell - can be used for any cell type.

– Partially sort cells by min value to find median cell.

– For each sub-tree, partially sort by max value to find median.

– Repeat . . .

Page 46: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

46EG2000

Kd-TreesKd-Trees

Searching Kd-Tree compare threshold (T) with min of root

cell, – If greater, then test cell at this node, then test

both children.

– if less need only progress down one half of tree.

Page 47: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

47EG2000

Performance - Lattice Sub-Division

Performance - Lattice Sub-Division

Shen, Hansen, Livnat & Johnson 1996. Uses the idea that each cell range can

be represented as a single point in the span space.

Subdivide span space into a lattice and place each data cell into a lattice element.

Page 48: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

48EG2000

Lattice Sub-DivisionLattice Sub-Division

Searching throws up 5 possible cases– case 1 - no intersection– case 2 - definite intersection– case 3 - test max only– case 4 - test min only– case 5 - test min and max

Page 49: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

49EG2000

Lattice Sub-DivisionLattice Sub-Division

In practice, lattice elements of equal size do not contain equal numbers of cells, so elements are allowed to be of unequal size.

This sub-division method suitable for use on parallel machines, elements are assigned to a different processor using a round-robin method.

Page 50: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

50EG2000

Performance - Interval TreesPerformance - Interval Trees

Cignoni, Marino, Montani, Puppo and Scopigno 1997.

Applies the optimally efficient Interval Trees data structure of Edelsbrunner to searching for active cells.– Find range for each cell and create a set

of ranges.– find the middle value of the overall

range of the data set, dr, and use it as the discriminating value for the root node.

Page 51: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

51EG2000

Interval TreesInterval Trees

– Using dr, at the root of the tree place all the intervals that contain dr, into two lists, one, AL, ordered by min, the other list, DR, ordered by max.

– For each of the 2 children, of the root create similar lists with respect to their discriminating value.

– Repeat . . .

Page 52: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

52EG2000

Interval TreesInterval Trees

To search the interval tree for a value T– if T < dr, scan list AL until list value >

T, use all these cells, then search left subtree only.

– if T > dr, scan list DR until list value < T, use these cells, then search right subtree only.

– If T = dr, just use cells in AL. This method has been described

in use with Structured data.

Page 53: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

53EG2000

Alternatives for visualization over the

network

Alternatives for visualization over the

network

Reduce size of surface to be sent to viewer by reducing number of triangles.

Can be done using mesh simplification methods.

Other alternatives exist . . .

Page 54: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

54EG2000

Web VisWeb Vis

Engel, Westermann & Ertl, Reduce triangle count by

creating triangle strips. Use a multi-resolution approach,

but rather than reducing the data by means of an error based model, take the user’s focus of interest and a given radius for hi quality mesh, the rest at lower quality.

Page 55: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

55EG2000

View Dependent

Isosurface

View Dependent

Isosurface

Livnat and Hansen 1998+– Only create isosurface from cells that will be

visible from user’s viewpoint, 3 step algorithm.

– Step 1 - find active cells by using Octree representation plus front to back traversal.

– Step 2 - coarse software visibility tests to further reduce cells used.

– Step 3 - send triangles to hardware.

Page 56: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

56EG2000

Surface SimplificationSurface Simplification

Discretized Marching Cubes - Montani et al 1994– Uses own lookup table– doesn’t interpolate along edges.

– combines co-planar triangles into larger polygons

Page 57: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

57EG2000

SlicingSlicing

Geometric intersection of a plane with the cells of the data set.

Simple approach is to test each cell against the plane to look for intersections.

Most cells are not required - hence performance improvements can be found by reducing number of cell tests.

Page 58: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

58EG2000

SlicingSlicing

Performance improvements easy for structured data when slice is axially aligned.

More difficult when cell has arbitrary orientation.

Page 59: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

59EG2000

SlicingSlicing

More difficult for unstructured data. Possible to use range based

methods from isosurfacing since problem can be reduced to a single range.– For axially aligned planes, construct

tree using appropriate x/y/z coordinate.– For arbitrary orientation, apply suitable

rotation then construct tree using appropriate x/y/z coordinate.

Page 60: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

60EG2000

Structure of TalkStructure of Talk

Data Enrich Map Render

Isosurface Slice Volume Render

Robustness

Accuracy

Topology

Shape

MarchingCubes

Surfacerep’n

Perform-ance

Structured

Unstructured

Page 61: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

61EG2000

Modelling the Data as Translucent Gel

Modelling the Data as Translucent Gel

Basic concept is to model data as a translucent gel

Classification step:

– maps data values to opacity via opacity transfer function

– .. and to colour via colour transfer function

Page 62: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

62EG2000

Classical Approach - Volume Rendering Integral

Classical Approach - Volume Rendering Integral

Cast rays through image plane into volume, and measure light received

– Kajiya and von Hertzen (1984)

– Max (1995) L

s

I = L0C(s)(s) exp[ -s

0 (t)dt ] ds

C(s)=light reflectedat point s

(s) = density atpoint s

lightdensity attenuation

imageplane

volume

Page 63: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

63EG2000

Simplifying the IntegralSimplifying the Integral

Approximate using Riemann sums (n = number of steps)

Approximate exponential by Taylor series and introduce opacity, , and unit spacing

Calculate recursively front-to-back as...

I = L0C(s)(s) exp[ -s

0 (t)dt ] ds

I = ni=0 C(is)(is)s

i-1j=0 exp [- (js)s]

I = ni=0 C(i)(i) i-1

j=0 (1 - (j))

Cout = Cin + (1-in)iCi

out = in + i(1 - in)

Compositing associative but not commutativeie can group but cannot re-order

{stop when = 1}

Page 64: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

64EG2000

The Two ApproachesThe Two Approaches

Image order:– from image to

volume– classical ray

casting method of Levoy (1988)

Recent attention:– integration– interpolation– different meshes– fast traversal

Object order:– from volume to

image– classical splatting

method of Westover (1989)

Recent attention:– better splatting– shear warp

rendering– different meshes– texture mapping – hardware advances

Page 65: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

65EG2000

Structure of TalkStructure of Talk

Data Enrich Map Render

Isosurface Slice Volume Render

Robustness

Accuracy

Topology

Shape

MarchingCubes

Surfacerep’n

Perform-ance

Structured

Unstructured

ImageOrder

ObjectOrder

Integration

Fast traversal

Interpolation

Meshes

Shear Warp

Splatting

Meshes

Textures

Hardw

are

Page 66: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

66EG2000

IntegrationIntegration

Riemann sums can be replaced by more accurate techniques (eg Simpson’s rule)

Recent work by Jung et al (1998) has found semi-analytical solution

– C(s), (s) expressed as 3rd degree polynomials via trilinear interpolation

– numerical approx to exponential term

I = L0C(s)(s) exp[ -s

0 (t)dt ] ds

cubicpolynomials

numericalintegration

Page 67: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

67EG2000

Maximum Intensity Projection

Maximum Intensity Projection

When performance rather than accuracy is the goal, we can avoid compositing altogether and approximate I by maximum intensity along ray

MIP : Maximum Intensity Projection

Often used in angiography...

Circle of Willis(University of Iowa)

Page 68: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

68EG2000

Maximum Intensity Projection

Maximum Intensity Projection

Performance is major issue

– lack of shading in image drives need for real-time rotation

– fast identification of maximum becomes important

imageplane

volume

- analytical maximumin each cell along ray- maximum of samplesalong ray- skip cells below maximum

.. but need to achievehigh quality tooSee Mroz et al (EG2000)for fast, high quality MIP

Page 69: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

69EG2000

InterpolationInterpolation

Sample points occur within cells, not at grid points, so we need to interpolate

Do we:– classify at grid

points, then interpolate

– interpolate, then classify

– ?

Classify - interpolate

– classification done as pre-processing

– smoothing effect can obscure detail

Interpolate - classify– classification now

within the inner loop of the ray cast (sample points are view dependent)

– in return, fine detail can be picked out

See Gasparakis (1999)

Page 70: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

70EG2000

Classify - InterpolateClassify - Interpolate

Wittenbrink et al (1998) point out a danger in interpolation after classification

Naïve colour interpolation would assign 3 parts yellow, 1 part blue to centre point…

… but if opacity of bottomleft is zero?

Correct approach is toweight according to opacity,so colour at centre is yellow!

=1 =1

=1=0

Page 71: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

71EG2000

Different MeshesDifferent Meshes

Fundamental problem is volume rendering compositing is non-commutative

Hence order matters Lesser problem for

rectilinear grid where cells are naturally ordered

Big complexity problem for curvilinear and unstructured grids

Page 72: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

72EG2000

Different Meshes - Curvilinear

Different Meshes - Curvilinear

Fruhauf (1994) algorithm transforms to rectilinear grid to ray cast, then back

Hong and Kaufman (1998) ray cast into curved volume

– find first cell, entry and exit point

– accumulate colour and opacity

– find next cell

Key is to find the sequence of cells efficiently

– cell faces projected to image plane

– bucket sort to get depth ordering

example of 3D complexity reduced to 2D problem

Page 73: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

73EG2000

Different Meshes - Unstructured

Different Meshes - Unstructured

Giertsen (1992) introduced idea of sweep plane

– Sweep plane contains all rays for 1 scan line

– Find cells intersecting plane and order (reduced to 2D problem)

– Keep set of active cells to exploit coherence

Silva and Mitchell (1997) lazy sweep ray casting algorithm

imageplane

intersectionwith tetrahedron

Page 74: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

74EG2000

Fast TraversalFast Traversal

Template-based ray traversal (Yagel et al, 1992)

Pre-process volume to identify regions of significance

– octree decomposition (Parker et al, 1999)

– boundary method

Boundary method– Wan et al (1999)– project boundary

cells to image plane

– create nearest and furthest buffers

– only process rays which intersect, and only process from nearest to farthest

Page 75: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

75EG2000

Structure of TalkStructure of Talk

Data Enrich Map Render

Isosurface Slice Volume Render

Robustness

Accuracy

Topology

Shape

MarchingCubes

Surfacerep’n

Perform-ance

Structured

Unstructured

ImageOrder

ObjectOrder

Integration

Fast traversal

Interpolation

Meshes

Shear Warp

Splatting

Meshes

Textures

Hardw

are

Page 76: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

76EG2000

Better SplattingBetter Splatting

Original splatting does shading then interpolate - causing smoothing effect

Recent work at Ohio has re-ordered to allow shading after interpolation, getting better detail

– Mueller et al (IEEE Vis99)

Page 77: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

77EG2000

Shear Warp RenderingShear Warp Rendering

To get fast traversal, shear volume by translating each slice… then can resample as shown

Project front-to-back to get intermediate image

Then warp image Note parallelised

versions

Page 78: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

78EG2000

Unstructured GridsUnstructured Grids

Cells are projected onto image plane

For all pixels covered by a cell, compositing operation applied

Ordering of cells is challenging computational geometry problem

Williams, Max and Stein (1998) describe high quality algorithm

Page 79: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

79EG2000

Texture MappingTexture Mapping

Exploits texture mapping and blending provided in OpenGL environments

– Volume is sliced parallel to viewing plane

– texture painted on to rectangle slices

– textured rectangles are composited

Page 80: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

80EG2000

Texture MappingTexture Mapping

SGI OpenGL Volumizer is software which exploits 3D texture mapping

Page 81: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

81EG2000

Texture MappingTexture Mapping

With 2D texture hardware, approach is still possible

Generate views parallel to co-ordinate planes

Choose closest to viewing direction

Example using VRML for medical volume visualization

Hendin, John, Schochet (1998)

Page 82: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

82EG2000

Hardware AdvancesHardware Advances

Holy grail: real-time volume rendering

Main searcher has been Kaufman through Cube architectures

Also major European effort

– VIZARD (Tubingen)

VolumePro System now commercially available from Mitsubishi’s RealTime Visualization

Page 83: 1 EG2000 Recent Advances in Visualization of Volumetric Data Ken Brodlie Jason Wood University of Leeds

83EG2000

ConclusionsConclusions

Volume Visualization– 1980s: basic algorithms– 1990s: enhancements in terms of

robustness, accuracy and performance

– 2000: hardware solutions Isosurface, slice or volume

render?– the winner is the user