2
** $Header: /cvs/projects/ogl-sample/main/gfx/lib/glu/libtess/alg-outline,v 1.1 2000/04/26 05:53:59 ljp Exp $
5
This is only a very brief overview. There is quite a bit of
6
additional documentation in the source code itself.
9
Goals of robust tesselation
10
---------------------------
12
The tesselation algorithm is fundamentally a 2D algorithm. We
13
initially project all data into a plane; our goal is to robustly
14
tesselate the projected data. The same topological tesselation is
15
then applied to the input data.
17
Topologically, the output should always be a tesselation. If the
18
input is even slightly non-planar, then some triangles will
19
necessarily be back-facing when viewed from some angles, but the goal
20
is to minimize this effect.
22
The algorithm needs some capability of cleaning up the input data as
23
well as the numerical errors in its own calculations. One way to do
24
this is to specify a tolerance as defined above, and clean up the
25
input and output during the line sweep process. At the very least,
26
the algorithm must handle coincident vertices, vertices incident to an
27
edge, and coincident edges.
30
Phases of the algorithm
31
-----------------------
33
1. Find the polygon normal N.
34
2. Project the vertex data onto a plane. It does not need to be
35
perpendicular to the normal, eg. we can project onto the plane
36
perpendicular to the coordinate axis whose dot product with N
38
3. Using a line-sweep algorithm, partition the plane into x-monotone
39
regions. Any vertical line intersects an x-monotone region in
41
4. Triangulate the x-monotone regions.
42
5. Group the triangles into strips and fans.
45
Finding the normal vector
46
-------------------------
48
A common way to find a polygon normal is to compute the signed area
49
when the polygon is projected along the three coordinate axes. We
50
can't do this, since contours can have zero area without being
51
degenerate (eg. a bowtie).
53
We fit a plane to the vertex data, ignoring how they are connected
54
into contours. Ideally this would be a least-squares fit; however for
55
our purpose the accuracy of the normal is not important. Instead we
56
find three vertices which are widely separated, and compute the normal
57
to the triangle they form. The vertices are chosen so that the
58
triangle has an area at least 1/sqrt(3) times the largest area of any
59
triangle formed using the input vertices.
61
The contours do affect the orientation of the normal; after computing
62
the normal, we check that the sum of the signed contour areas is
63
non-negative, and reverse the normal if necessary.
66
Projecting the vertices
67
-----------------------
69
We project the vertices onto a plane perpendicular to one of the three
70
coordinate axes. This helps numerical accuracy by removing a
71
transformation step between the original input data and the data
72
processed by the algorithm. The projection also compresses the input
73
data; the 2D distance between vertices after projection may be smaller
74
than the original 2D distance. However by choosing the coordinate
75
axis whose dot product with the normal is greatest, the compression
76
factor is at most 1/sqrt(3).
78
Even though the *accuracy* of the normal is not that important (since
79
we are projecting perpendicular to a coordinate axis anyway), the
80
*robustness* of the computation is important. For example, if there
81
are many vertices which lie almost along a line, and one vertex V
82
which is well-separated from the line, then our normal computation
83
should involve V otherwise the results will be garbage.
85
The advantage of projecting perpendicular to the polygon normal is
86
that computed intersection points will be as close as possible to
87
their ideal locations. To get this behavior, define TRUE_PROJECT.
93
There are three data structures: the mesh, the event queue, and the
96
The mesh is a "quad-edge" data structure which records the topology of
97
the current decomposition; for details see the include file "mesh.h".
99
The event queue simply holds all vertices (both original and computed
100
ones), organized so that we can quickly extract the vertex with the
101
minimum x-coord (and among those, the one with the minimum y-coord).
103
The edge dictionary describes the current intersection of the sweep
104
line with the regions of the polygon. This is just an ordering of the
105
edges which intersect the sweep line, sorted by their current order of
106
intersection. For each pair of edges, we store some information about
107
the monotone region between them -- these are call "active regions"
108
(since they are crossed by the current sweep line).
110
The basic algorithm is to sweep from left to right, processing each
111
vertex. The processed portion of the mesh (left of the sweep line) is
112
a planar decomposition. As we cross each vertex, we update the mesh
113
and the edge dictionary, then we check any newly adjacent pairs of
114
edges to see if they intersect.
116
A vertex can have any number of edges. Vertices with many edges can
117
be created as vertices are merged and intersection points are
118
computed. For unprocessed vertices (right of the sweep line), these
119
edges are in no particular order around the vertex; for processed
120
vertices, the topological ordering should match the geometric ordering.
122
The vertex processing happens in two phases: first we process are the
123
left-going edges (all these edges are currently in the edge
124
dictionary). This involves:
126
- deleting the left-going edges from the dictionary;
127
- relinking the mesh if necessary, so that the order of these edges around
128
the event vertex matches the order in the dictionary;
129
- marking any terminated regions (regions which lie between two left-going
130
edges) as either "inside" or "outside" according to their winding number.
132
When there are no left-going edges, and the event vertex is in an
133
"interior" region, we need to add an edge (to split the region into
134
monotone pieces). To do this we simply join the event vertex to the
135
rightmost left endpoint of the upper or lower edge of the containing
138
Then we process the right-going edges. This involves:
140
- inserting the edges in the edge dictionary;
141
- computing the winding number of any newly created active regions.
142
We can compute this incrementally using the winding of each edge
143
that we cross as we walk through the dictionary.
144
- relinking the mesh if necessary, so that the order of these edges around
145
the event vertex matches the order in the dictionary;
146
- checking any newly adjacent edges for intersection and/or merging.
148
If there are no right-going edges, again we need to add one to split
149
the containing region into monotone pieces. In our case it is most
150
convenient to add an edge to the leftmost right endpoint of either
151
containing edge; however we may need to change this later (see the
158
These are the most important invariants maintained during the sweep.
159
We define a function VertLeq(v1,v2) which defines the order in which
160
vertices cross the sweep line, and a function EdgeLeq(e1,e2; loc)
161
which says whether e1 is below e2 at the sweep event location "loc".
162
This function is defined only at sweep event locations which lie
163
between the rightmost left endpoint of {e1,e2}, and the leftmost right
166
Invariants for the Edge Dictionary.
168
- Each pair of adjacent edges e2=Succ(e1) satisfies EdgeLeq(e1,e2)
169
at any valid location of the sweep event.
170
- If EdgeLeq(e2,e1) as well (at any valid sweep event), then e1 and e2
171
share a common endpoint.
172
- For each e in the dictionary, e->Dst has been processed but not e->Org.
173
- Each edge e satisfies VertLeq(e->Dst,event) && VertLeq(event,e->Org)
174
where "event" is the current sweep line event.
175
- No edge e has zero length.
176
- No two edges have identical left and right endpoints.
178
Invariants for the Mesh (the processed portion).
180
- The portion of the mesh left of the sweep line is a planar graph,
181
ie. there is *some* way to embed it in the plane.
182
- No processed edge has zero length.
183
- No two processed vertices have identical coordinates.
184
- Each "inside" region is monotone, ie. can be broken into two chains
185
of monotonically increasing vertices according to VertLeq(v1,v2)
186
- a non-invariant: these chains may intersect (slightly) due to
187
numerical errors, but this does not affect the algorithm's operation.
189
Invariants for the Sweep.
191
- If a vertex has any left-going edges, then these must be in the edge
192
dictionary at the time the vertex is processed.
193
- If an edge is marked "fixUpperEdge" (it is a temporary edge introduced
194
by ConnectRightVertex), then it is the only right-going edge from
195
its associated vertex. (This says that these edges exist only
196
when it is necessary.)
202
The key to the robustness of the algorithm is maintaining the
203
invariants above, especially the correct ordering of the edge
204
dictionary. We achieve this by:
206
1. Writing the numerical computations for maximum precision rather
209
2. Making no assumptions at all about the results of the edge
210
intersection calculations -- for sufficiently degenerate inputs,
211
the computed location is not much better than a random number.
213
3. When numerical errors violate the invariants, restore them
214
by making *topological* changes when necessary (ie. relinking
218
Triangulation and Grouping
219
--------------------------
221
We finish the line sweep before doing any triangulation. This is
222
because even after a monotone region is complete, there can be further
223
changes to its vertex data because of further vertex merging.
225
After triangulating all monotone regions, we want to group the
226
triangles into fans and strips. We do this using a greedy approach.
227
The triangulation itself is not optimized to reduce the number of
228
primitives; we just try to get a reasonable decomposition of the
229
computed triangulation.