Computer Based
Terrain Visualisation Techniques

"From Points to Wood"

Written by Paul Bourke
August 1988


Introduction

This report describes some of the computer based techniques which can be used to aid in the visualisation of terrain. In the following discussion it is assumed that we have acquired data in a suitable form which allows us to define the surface on a computer. This usually starts as a series of points lying on the surface and stored as x,y,z triples. There are many ways of getting this data, the example here was digitised from contour maps using a large scale digitising tablet connected to a computer. The next step is normally to turn the digitised data into a series of polygonally bounded planes, called facets, so that they can be read, viewed and manipulated by 3D modelling and rendering software. The point data is this example is transformed into a regular polygonal mesh by an algorithm called Delaunay triangulation See an earlier discussion by myself for details of this algorithm and computer program which transforms randomly distributed spot heights into triangular or regular meshes and exports this surface description in a format suitable for most 3D modelling and/or rendering programs.

Contouring

Contouring has been the traditional way of presenting 3D terrain data on 2D sheets of paper. It has advantages of having units (contour values) so precise height calculations can be made. Contour lines are usually drawn with line segments so they can easily be transferred to large scale hard copy devices such as plotters. The main problem with this visualisation technique is that it does not give a good 3D impression of the terrain, the best most people can determine is that some localised parts of the surface are higher or lower than others.

Mesh representations

This is the most straightforward way of rendering the 3D data, it is a direct perspective viewing transformation of the computer database. (Almost any 3D modelling package can display, view, and print this representation, MicroStation was used here due to its advantages when handling very large geometric databases).


Figure 1
 

Figure 1 shows the example landscape represented in this way. The reasons why this makes the landscape look three dimensional are: the automatic shading of the steep parts of the terrain because the mesh lines appear closer together, the flat areas are obvious due to the regularity of the mesh, the perspective transformation gives a simple depth cueing due to the mesh lines becoming denser. These effects generally require that the mesh density approaches that of the viewing device, or at least that the projection of the mesh onto the viewing plane is close to the resolution of the device.

This rendering can also be in colour. The colour may be related to height but it could also be some other attribute such as ground cover, population, etc.

The example in Figure 1 is a 128x128 cell mesh and it consists of about 33,000 line segments. On many high end computer platforms this can be drawn in real time so that the user can "fly" about the landscape. This number of polygons can however cripple many desktop machines. For this reason good terrain modelling software allows the user to view the surface at a range of resolutions.

Vertical contouring

Figure 2
 

While horizontal contouring has long been a popular method, vertical contours can also be effective. The reasons for the success of this rendition method are the same as for mesh representations, it generates dark (dense) areas which would naturally be shadowed in direct sun light.

Figure 2 is an example of such a rendering of our demonstration landscape. This rendering was created with a 3D modelling package called Vision3D written by myself. There is a considerable reduction in the data needed to draw this form of rendering, unfortunately the effect is only satisfactory when the contour lines are about 45 degrees to the view direction and so animations are not usually possible.

Shaded renderings

Higher levels of realism can be achieved by simulating more closely how the terrain would appear in reality. There are a wide range of techniques for accomplishing this, each technique generally involves a trade off between realism and computation time.


Figure 3
 

Figure 3 shows a relatively sophisticated rendering This rendering was created using a raytracing/radiosity package called Radiance. It was chosen mainly because it was available on a very fast hardware platform. One undesirable characteristic of this technique can be seen in the patches visible on the surface. These are the result of the limits in the data available in the description of the surface. The courser the mesh the more obvious the patches are. The finer the mesh the more data the rendering process must handle and the longer it will take.

There are techniques (Phong shaded polygons) which can reduce and even remove the patchiness of the surface by interpolating the surface normals between the facets. These techniques however require more computation time and have the undesirable characteristic that the shadows are still based on the gridded geometry and do not appear smooth. The example in Figure 3 uses the same data density as the mesh shown in Figure 1.

Physical models

Given the digital model it is possible to determine the path required to manufacture the surface using a computer controlled milling machine. In its simplest form this is a drill which can be controlled backward and forward and up and down over a piece of wood say, so as to cut away unwanted portions.


Figure 4

Figure 4 shows the result of milling our example landscape from wood. I used a OH-FANUC, model 2R-NC milling machine operated by the School of Engineering, Auckland University. The machine is controlled by the MasterCAM software For this example the wood is about 300mm square and the drill bit used in about 5mm radius. Note that it is not necessary to use a particularly fine bit because the wood is cut away at the edge of the bit not from the bottom. The bit size then only determines the narrowest valleys that are possible. This physical scale model of the landscape has big advantages for visualisation purposes. The model is to scale, although in this case there is a 2 times exaggeration in height. The viewer can instantaneously view the terrain from various positions and angles by simply turning the model about. Tactile exploration of the model is of course possible and can be informative as well as satisfying.

An extension of this technique would be to automatically draw features such as roads, boundaries, contours, etc onto the landscape. This could be done with a robot arm holding a pen or with a laser which would burn the line features onto the surface of the wood.


Data reduction in terrain modelling

Written by Paul Bourke
June 1994


A common characteristic of terrain modelling/rendering exercises is the vast amount of data involved. This is particularly severe where the spot height information is a result of some automated survey or when it results from digitised contours. In addition there are a number of activities which require very different levels of detail, for example the detail required for high quality renderings may be quite high. The information that a 3D modelling package can handle while retaining user interaction may be quite low. The need for multiple terrain representations can arise naturally in modelling applications, for example it may be necessary to acquire a more course terrain model for interactive modelling and substitute a more detailed version for final presentation. As with many computer based activities there is often a gap between what one might like to and what is practical given disk, memory, or processing resources.

There are two common representations for computer based terrain models. The first consists of a triangular mesh of polygons, this is normally the result from a triangulation process. One advantage of this approach is that it naturally generates detail in the regions that are sampled more frequently, these are normally the regions which have more height variation. This representation is a more faithful representation of the underlying data since the sample points are generally the vertices of the triangular mesh.

The second form of representation is as a regular mesh of rectangular polygons, these are generated by estimating the heights at each corner of the mesh cells. The main advantage with this is that multiple resolution meshes can be readily generated from the same dataset of spot heights. This form is tend to provide more 3D visual cues as the wire frame rendered mesh acts a type of slope based shading.

Triangulation example

In this example the original digitised aerial survey resulted in 600,000 spot heights, covering a rectangular region of 42km by 64km. Converting this directly to a triangular mesh would result in a surface of approximately twice that number of polygons, 1.2million!. This number of polygons would certainly be a handful for most software packages especially if interactive 3D control was necessary.

One approach to data reduction is to filter the spot heights file so as to evenly distribute the points over the region. This is done by checking the distances between all points so that no two points are closer than some user chosen distance. An improved approach would be to incorporate the height variation between neighbouring points so that highly sampled areas are better represented, this assumes that regions are densely sampled because of increased height variation.

Below are some examples of the surface and the number of resulting polygons for 1,3,and 5km filter distances.

5km radius, 116 polygons

3km radius, 316 polygons

1km radius, 2600 polygons

Notice there are only a few polygons in the lake regions where no sampling was necessary.

Gridding example

This example uses another terrain database containing 100,000 points over a 650m square area. These were used to produce a gridded representation at a user selected resolution and thus file size. This very precise control of resolution is not possible in the previous filtering process, the user didn't know the resulting number of polygons beforehand. The exact size of the grid cells is also precisely known which can be useful for rough distance estimates.

The following are some examples of 10, 20, and 65 square meshes showing the number of polygons involved.

10x10 grid, 65m grid size, 40 polygons

20x20, 32.5m grid size, 200 polygons

65x65, 1m grid, 1850 polygons

There is normally a further overhead if a rectangular mesh surface is to be used for rendering purposes because the 4 point facets above will generally not be coplanar. The simplest solution is to split each facet into two along the diagonal thus increasing the number of polygons by a factor of 2. A better method is to interpolate the midpoint, unfortunately this increases the number of polygons by a factor of 4.

In summary, the two representations can be used to create a terrain model of any desired data size. Indeed one would normally generate a number of surfaces with different resolution, the particular one used at any given time would depend on the applications ability, the response times, and quality/precision goals.


Terrain morphing

Written by Paul Bourke
April 2001

Source code that created the examples below.
QuickTime animation of example 3: morph.mov.gz (1.2MB)


Morphing is a popular technique in the computer graphics industry extensively used in Michael Jackson videos but also used in countless movies and television ads. The technique was first used in the early 1980's, one of the first movies that used morphing was "Indiana Jones and the last Crusade". Morphing is a very different technique to fading, in that case the colour of each pixel in an intermediate frame is a linear interpolation between the colour of the corresponding pixel in the start and stop frame.

For example
colourintermediate = colourstart + mu (colourstop - colourstart)

In morphing one tries to capture a sense of the geometric transformations between the start and stop frame. In what follows a standard morphing technique as presented by Thaddeus Beier at the 1992 Siggraph will be applied to create smooth transitions between two terrain datasets. For example the application in mind was the morphing of data from continental drift simulations. The technique and algorithm here applies equally well to the more traditional image morphing. The strict term for this method is "field morphing" because the operator chooses related regions in the start and stop frames, during the morph the relationships between these regions is preserved.

In order to understand the morph transformation, assume we have a start and stop surface. On each of these surfaces we create a single directed line, this pair of lines indicates a transformation relationship between the two surfaces. For example, in the following the intention is that the region around the line P1 to P2 is rotated and shrunk during the morph between the two frames.

Now consider estimating the surface height at a point P on an intermediate surface. A directed line P1->P2 on this intermediate surface is formed by interpolating between the lines in the start and stop image. In order to estimate the value at point P compute u and v, these are the normalised distance along the lines P1->P2 and P1->Perp respectively. Note that there are two possible perpendicular vectors, in what follows it doesn't matter which perpendicular is used as long as a consistent choice is made.

P = P1 + u (P2 - P1) + v (Perp - p1)

These values of u and v are applied the line in the start and stop surface to get the corresponding point in those surfaces. The line on the intermediate image is a linear interpolation between the corresponding lines in the start and stop surface. The surface height in the intermediate surface is a cross fade of these two height samples. If the animation sequence is controlled by a parameter mu that ranges from 0 (start) to 1 (stop) and the two height samples are z1 (start) and z2 (stop) then the intermediate height is as follows.

z = z1 (1 - mu) + z2 mu

Most morphs will require more than a single line to define the related regions in the start and stop surfaces. For every point on the intermediate surface each line gives a point on the start surface, a weighted sum of all these points is used to choose the point to sample on the start surface. The same applies for the stop surface. The cross fade is then applied to the weighted sum estimate of the corresponding points on the start and stop surface.

The weight is predominantly determined by the distance of the point and the line on the control surface. This is intuitively obvious, the influence a line plays on a point is inversely related to the distance of the point to the line. Different forms of the weighting can change the exact details of the field on influence, for the details see the sample code.

Examples

In the examples that follow, fractal landscapes are created within the red outlines. The control lines are shown in green except for example 2 where the control lines are identical to the outline. Samples from an animation between the start and stop surfaces are given for both simple blending and morphing. While the blend is a valid smooth transition between the two surfaces, the sample below should convey the richer transition given by the morph.

Example 1 - Rotation

This first example illustrates rotation, note the single control line in the last frame is at 90 degrees to where it is in the first frame. The blend case doesn't behave too objectionably and may be the desired effect in some application. In the morph case there is a clear rotation of the island features.

Start frame
Stop frame

Blend

Morph

Example 2 - Translation

This is an example of translation, as with the above example there is a stunning difference between the morph case and the straightforward blend of the height fields. In the morph case the island slowly moves across the ocean while in the blend one island drops away while the other appears.

Start frame
Stop frame

Blend

Morph

Example 3 - Separation

This is a more complicated example, the main land mass rotates and a small island splits off. In the blend example the small island appears but in the morph animation the land mass clearly rotates while the island slowly pulls away.

Start frame
Stop frame

Blend

Morph

Example 4 - Lake Creation

This example is only fully appreciated as an animation sequence, the left and right arms are drawn around and closes off the gap. This example does illustrate the difficulty imposed by the requirement of having the same number of control lines in the start and stop surface.

Start frame
Stop frame

Morph

References

Digital Image Processing of Earth Observation Sensor Data. Bernstein, R. IBM J. Res. Development, 20:40-57, 1976.

Feature-Based Image Metamophosis. Beier, T. Siggraph 1992

Space Deformation Models Survey. Bechmann, D. Computers & Graphics, 18(4):571-586, 1994.

Three-Dimensional Distance Field Metamorphosis. Cohen-Or, D., Levin, D., Solomovici, A.. Transactions on Graphics, 1998.

Extended free-form deformations: A sculpturing tool for 3D geometric modeling. Coquillart, S. Computer Graphics (SIGGRAPH '90 Proceedings), volume 24, pages 187--196, 1990.

Conformal image warping. Frederick, C. and Schwartz, E.L. IEEE Computer Graphics and Applications, 10(3):54-61, March 1990.

Establishing Correspondence by Topological Merging, A New Approach to 3D Shape Transformation. Kent, J., Parent, R., Carlson, W.

The Morphological Cross-Dissolve. Novins, K., Arvo, J. Conference Abstracts and Applications, page 257. ACM SIGGRAPH, August 1999.

A Morphable Model for the Synthesis of 3D Faces. Blanz, V., Vetter, T. Proceedings of SIGGRAPH 99, pages 187-194. ACM SIGGRAPH, August 1999.

Digital Image Warping. Wolberg, G. IEEE Computer Society Press, 1990


Terragen Animation Extension (tgs_interp)

Written by Paul Bourke
July 2003, updated November 2004

"Terragen is a scenery generator, created with the goal of generating photorealistic landscape images and animations.....capable of near-photorealistic results for professional landscape visualisation, special effects, art and recreation." [Terragen web site].

Motivation

There are a number of environments in which one might want to create content using Terragen but for which Terragen doesn't provide explicit support eg: stereoscopic projection, fisheye frames for planetarium domes, panoramic images. The following describes a utility that solves this problem and as a side effect conforms to conventions that allow time consuming animation rendering to be performed locally on a large rendering cluster of 200+ processors.

The utility is called "tgs_interp" mostly because it provides keyframe interpolation (supported by keyframe files in Terragen called a .tgs file). The basic work flow is to create a tgs file using Terragen or plug-ins, use tgs_interp to read this tgs file and instruct it to create a number of other tgs files which will then be read and rendered by Terragen. In the case of stereoscopic projection two tgs frames result from each interpolated frame. For cubic projection 5 or 6 tgs frames result for each interpolated frame. The utility is written for UNIX command line operation, currently compiled for Mac OS-X and Linux.

Usage: tgs_interp [options] tgsfile
Options
   -n n   number of tween frames (default: 1)
   -c n   calculate cubic views, 5 or 6 (default: off)
   -s n   calculate stereo pairs, supply eye separation (default: off)
   -p n   add a pitch offset in degrees, positive pitch forward (default: 0)
   -f n   tgs splitting mode, 0=single, 1=lots, 2=stereo, 6=cubic
   -z n   adjust camera height (default: 0)
Projection options

The two environments not directly supported by Terragen and addressed by this utility are:

  • Stereoscopic (-s)
    In this case the camera is offset to the left and right of the non-stereoscopic position. The width of the left and right images should be widened for rendering in Terragen and truncated afterwards to create the correct stereo pairs, this is discussed in more detail here and later in this document.

  • Cubic (-c).
    Cubic renderings are created by rendering 5 or 6 perspective frustums, namely the frustums formed from the center of a cube about the camera and each of the faces of the cube. These renderings can then combined into a fisheye image (or other projections such as panoramic or spherical) using techniques described here. Note that the QuickTime VR support in Terragen isn't quite good enough for creating fisheye animations because the faces of the cube in that case are aligned with the coordinate axes rather than the camera view direction.

Frame output options (-f)

This option determines the format of the output tgs file(s). The options are as follows.

  • 0 - All the frames are written to a single tgs file, the default. This is usually the choice when tgs_interp is simply used as a keyframe interpolation.

  • 1 - Each frame is written to a single tgs file. This is intended primarily for locally developed cluster based distributed rendering software, use with care...it can create lots of files.

  • 2 - This is intended for use with stereoscopic output, the left and right eye frames will each be written to their own tgs file. Recommended for stereoscopic output.

  • 3 - This is intended for cubic output, each of the 5 or 6 faces (left, front, right, back, top, bottom) will be written to their own tgs file. Recommended for cubic output.

Additional options (-z and -p)
  • Height adjustment (-z)
    This raises or lowers the camera along the camera "up" vector.

  • Pitch adjustment (-p)
    This will pitch the camera forward or upward by some angle. Primarily intended for creating fisheye content for planetarium domes which have different ideal horizon levels. Note that this can also be performed using the software that converts cubic views to a fisheye image.

Notes on stereoscopic rendering

There are a number of ways of formulating and setting up stereoscopic rendering, the approach taken here will be the correct parallel camera method and since Terragen doesn't support offaxis frustums it will be based upon the techniques described here. In essence this involves rendering frustums that are slightly wider than desired and truncating the images to give the correct stereo pairs. The following example outlines how the author would set up a stereoscopic rendering using Terragen and tgs_interp.

  • The horizontal camera aperture "a" (also called the horizontal field of view) is controlled in Terragen by the zoom setting. The relationship is

    zoom = 1 / tan(a/2)

    So in this example if the intended horizontal aperture is 60 degrees, the Terragen zoom is 1.732

  • The degree to which the image width (as specified in Terragen) should be increased is given by

    delta = e w / (2 fo tan(a/2))

    So in this example, if the final image width "w" is to be 1024 pixels and the eye separation 100 units and the distance to zero parallax is 3000 units, then amount by which the image width will be extended is 30 pixels. Note this is based upon the eye separation being 1/30 of the distance to zero parallax, a common ratio. It is often more useful to think in terms of the zero parallax distance and then use 1/30 of that for the eye separation.
    So, Terragen should be instructed to render an image that is 1054 pixels wide. After the images are rendered 30 pixels will be removed from the left of the left eye image and 30 pixels will be removed from the right of the right eye image. This process in effect creates offaxis frustums even though only symmetric frustums were available.

  • And finally, to be strictly correct the aperture and hence the zoom setting in Terragen needs to be recalculated, the expression is

    2 atan((w + delta) tan(a/2) / w)

    In this example the intended aperture is 61.443 degrees and the zoom to be set in Terragen is 1.683. This is a relatively minor adjustment that is often not bothered with.


Left eye, model courtesy of Peter Morse
 
Right eye

Notes on cubic rendering and derived projections

The width and height of the rendered image should be set equal before rendering. The zoom factor set by the user will be reset to 1, in other words a horizontal (and vertical) aperture of 90 degrees.


Unfolded cube, model courtesy of Peter Morse


Derived spherical panoramic


Derived cylindrical panoramic


Derived fisheye

Further notes

  • The Mac and MSWindows version of Terragen seem to be slightly out of step with how animation files are named. Until this is resolved tgs_anim will use it's own conventions that just happen to be those preferred by the author and most compatible with local conventions used for stereoscopic content and animations for planetariums.

  • This utility has been tested/checked with the current version of Terragen and the format/style of the tgs files it creates. Obviously future versions of Terragen may change details of the tgs files that this utility assumes, or new features may be added for which this utility is ignorant. Every attempt will be made to keep this utility in step with the current version of Terragen.