Approximation of Texture Coordinates

by Johan "Yomat/Nature" Mattsson

This document is closely related to texture-mapping of polygons. Thus you need to know some about drawing texturemapped polygons before you will have any use for the information here. In texture-mapping we are assuming that we already know the texture coordinates, associated with each vertex of the polygon. Now we want a way to do a more object related study of texture mapping, where we want to apply a texture surface to a 3D object. A 3D object that we consider being composed of many polygons; a so called polygon mesh.

A texture image can be tiled over an object or a polygon several times. In effect this is like putting an infinite amount of textures side by side in every direction while keeping the same relation to itself in all sub-images.
Number-wise, the original texture image is placed with its upper left corner at the origin. The lower right corner is at the coordinate (1.0, 1.0). This means that a polygon with (u,v)-coordinates: ((2.3, 4.3), (2.5, 5.4), (3.4, 4.6)) will be mapped exactly the same way as a polygon with coordinates  ((0.3, 0.3), (0.5, 1.4), (1.4, 0.6)). The latter case is the closest to the origin and that we call the normalized tiling case. See also the figure below.

Surface description
Information about surfaces, attributes and mapping types partly comes from developer documents for the LightWave3D object format. I have also extracted some information from buggy sources released by NewTek.

Each face is associated with a surface. A texture map surface has an image reference and the following parameters associated with it:

  • Center, is the relative center of the texture coordinates to the object 3D-coordinates.
  • Size, denotes the scale factor of the texture coordinate to the object 3D-coordinates.
  • Flags, tells us the direction of the image relative to the object coordinates.
  • Tiling, gives the number of times an image is to tiled over a spherical or cylindrical mapped object. There is one parameter for width and one for height. Width is around the defined mapping axis and height is alongside the same axis.
We have three different mapping types that each surface can have. There are plenty of documents and books that describes the visual effect and how they are to be used in an aesthetic manner. We satisfy with describing the procedural aspects. All examples are when the mapping axis is set to be the y-axis, and all coordinates are centered as follows:
x = x - center.x;
y = y - center.y;
z = z - center.z;
  • Planar: This mapping is the straightforward one. A vertex coordinate (x,y,z) will simply become the (u,v)-coordinate. Since (x,y,z) is three dimensions and (u,v) is two, we have three different axis to use. The choice of axis is specified in the flag entry.
u = x/size.x + 0.5;
v = -z/size.z + 0.5;
  • Cubic: Similar to planar mapping. The axis of the mapping is given by the direction of the polygon instead of predefined. The axis chosen, is the one which direction vector is most close to the normal of the face of the polygon.
  • Spherical: The (u,v)-coordinates is simply the conversion from Euclidean coordinates (x, y, z) to spherical coordinates. See figure below. Again, the polar axis location depends on the axis flag.
longitude = atan2((float)-x, z);
latitude = atan(y / sqrt(x*x + z*z));

longitude = 1.0 - longitude / (2*PI);
latitude = fabs(0.5 - latitude / PI);

u = longitude - floor(longitude);
v = latitude;
  • Cylindrical: Straightforward conversion from 3D coordinates to cylindrical coordinates. Practically it is a degenerate case of spherical mapping. We use only the circular of the polar coordinates as u and do a planar mapping for v.

Actual Approximation

The previous section showed how we from a 3D-coordinate, of an objects polygon, can calculate an (u,v)-coordinate. If we want to draw the texture-mapped polygon, we can calculate the (x,y,z)-coordinate on the polygon for every pixel. Then we also have to do the mapping calculation for every pixel. This method works perfectly. However, this approach has little to do with the interpolation techniques we want to use when texturemapping a polygon. Secondly, this method would be hideously slow; we are considering a function call with 10-20 FPU-instructions and several calls to arctangent functions.

If we satisfy ourselves with calculating the (u,v) for only the vertices of the polygon, we will get a number of problems connected with the interpolation mapping techniques:

When we are doing planar and cubic mapping, we can get what we call a wrapping-problem. This occurs when the texture image seam (the line formed where the image meets the copy of itself) crosses an edge of a polygon. Similarly; when any of the resulting (u,v)-coordinates is larger or smaller than the image boundaries. Here we consider two possible ways of solving this problem.

The straightforward way is to split all polygons where they cross the image seam. It works nicely, but have a major drawback. Splitting might result in many more polygons. The more polygons, the slower the handling and drawing of the object will be.

Another approach, and the one we chose to use, is to produce a new texture image. This new image contains the problematic image tiled as much as necessary, so that all faces can be mapped without a seam problem. We also want to minimize the size of the larger generated image. It can be done by the following steps:

  1. Find the maximum and minimum of both u and v coordinates of a polygon.
  2. Calculate the displacement coordinates for the polygon; the values that normalizes the coordinates to the origin of the image.
  3. /*
      This is the optimizing step and it also removes all negative coordinates.
    if (minu < 0) 
      udisp = ceil(-minu);
      udisp = -floor(minu);
    if (minv < 0) 
      vdisp = ceil(-minv);
      vdisp = -floor(minv);
  4. Add the displacement to all (u,v)-coordinates of the polygon.
  5. During the polygon handling we find the greatest u and v coordinates for all polygons associated with the image.
  6. At the image handling stage, we use the greatest u and v to deduce the size and amount of tiles for the new texture image.
When a texture is tiled many times a very large image will be generated with this method. Naturallt we want to limit the image sizes to the common 256x256 texels. We also make use of modulo 256 with logical AND operators.Thus it is enough if we limit the image generation to a maximum of 256x256, since it will be tiled automatically by the polygon routine.

A keen observation will tell that we might get incorrect result, if texture image will not fit a whole amount of times into the 256 wide memory buffer. To avoid such problems when doing much tiling of a texture onto an object, the User Designer should only use images that have sizes that
are a power of two (256,128,64 etc).

The use of spherical and cylindrical mapping, adds interesting features to the modeling stage of a 3D world. Unfortunately it introduces nasty problems when it comes to finding the appropriate (u,v)-coordinates of the vertices.

The main problem with these mapping types is that they will not work or look proper when not used on the right kind of objects. The approximation techniques below, tries to handle the case when the object we are trying to map has a similarity to the geometry of its mapping type. For instance, trying to spherical map a cube might produce very bad results, but a sphere or a cone will look better.

The circular coordinates of both mapping types introduces a new kind of wrapping problem. Since we are dealing with angles, we can only get 0.0 to 1.0 (360 degrees) as u-coordinate. The bad thing happens when a the image seam crosses a polygon so that we get angles close to 1.0 and 0.0 at the same time, where we should get more than 1.0 (or similarly; negative coordinates). It is clear that we can not apply the tiling trick discussed above directly to this problem. We have to have a method to decide when a polygon is crossing a seam, and then do necessary modifications so that the tiling will be possible.

  •  Fact: The difference between two circular coordinates (angles) can not be more than 0.5 (180 degrees).
Using this fact it is possible to compare the difference between all circular coordinates (u-coordinates) of the polygon. If any of the differences is larger than or equal to 0.5, that polygon is crossing a texture seam.
for (j=0; j < face->count; j++)
  for (k=0; k < face->count; k++)
    if (0.5 <= fabs(uv[j].u - uv[k].u))
      flag |= 1;
If there is such a difference we have to add 1.0 to all u-coordinates, of the polygon, that are less than 0.5.

Consider the relation concerning distances between the 3D coordinates, of the vertices of the polygon, and the mapping axis. If the distances are relatively larger than a certain limit, the coordinate closest to the axis will loose its meaning in relation to the u-coordinates of the other vertices. See Figure below. The constant limit value with the best results, has to be discovered by experimenting.

The best method to solve this problem would be to split all polygons that has vertices too close to the axis. The split should be done in a manner such that, the difference in angles should be no more than a certain threshold. This would result in a large amount of triangle slivers, all with the small pointing towards the axis.

In real-time applications the splitting method is not a good option. The quite approximative solution suggested, is to assume that all problematic polygons have only one vertex close to the axis. The other vertices of the polygon are assumed to have relatively small angle-distance and large relative distance to the axis. If this assumption holds. We can ignore the u-coordinate of the vertex that is close to the axis. Instead we calculate a new u-coordinate by averaging the u-coordinates of the other vertices.

For this to work we must have a way to deduce which vertices that are in fact too close to the axis. A simple way that we chose; check if a vertex coordinate, in the plane orthogonal to the axis, is under a certain predefined limit. The following code does the trick:

#define TEX_EPSILON 0.05

  float uval = 0.0;
  for (j=0; j < face->count; j++)
      vrt = &obj->verts[face->vx[j]];
      /* Different tests for different mapping axis. */
      if ((surf->tflags & SRF_TFLAGS_X
           && fabs(vrt->y) < TEX_EPSILON
           && fabs(vrt->z) < TEX_EPSILON)
          (surf->tflags & SRF_TFLAGS_Y
           && fabs(vrt->x) < TEX_EPSILON
           && fabs(vrt->z) < TEX_EPSILON)
          (surf->tflags & SRF_TFLAGS_Z
           && fabs(vrt->x) < TEX_EPSILON
           && fabs(vrt->y) < TEX_EPSILON))
          index = j;
        uval += uv[j].u;
  if (index >= 0)
    uv[index].u = uval/(face->count-1);

After enough experimenting a good TEX_EPSILON limit can be found. For most of the cylindrical and spherical mapped objects that has been considered, the results are visually much more appeasing than just to ignore this special case.

Conclusion and References

With some thoughts its easy to realise that its impossible to approximate the coordinates pefectly and sometimes not even close. However one can achieve good enough results as can be seen in some of the demo-productions out there like 'Rise' and 'Relic'  which doesent have the standard planar type mapping everywhere.

One quick practical observation is that for a renderer to make use of the texture coordinates of the scheme described above, it need to have support for distinct texture-coordinates for each polygon. This compared to, for example, planar demostyle environment-mapping where one uv-coordinate per vertex is enough.

The authors version of 'uvcode'


(c) 1/2000 by Johan Mattsson (email: d92jma at