Line Geometry for 3D Shape Understanding

and Reconstruction

Helmut Pottmann, Michael Hofer, Boris Odehnal, and Johannes Wallner

Technische Universität Wien, A 1040 Wien, Austria.

{pottmann,hofer,odehnal,wallner}@geometrie.tuwien.ac.at

Abstract. We understand and reconstruct special surfaces from 3D

data with line geometry methods. Based on estimated surface normals we
use approximation techniques in line space to recognize and reconstruct
rotational, helical, developable and other surfaces, which are character-
ized by the configuration of locally intersecting surface normals. For the
computational solution we use a modified version of the Klein model of
line space. Obvious applications of these methods lie in Reverse Engi-
neering. We have tested our algorithms on real world data obtained from
objects as antique pottery, gear wheels, and a surface of the ankle joint.

Introduction The geometric viewpoint turned out to be highly successful in

dealing with a variety of problems in Computer Vision (see, e.g., [3, 6, 9, 15]).
So far mainly methods of analytic geometry (projective, affine and Euclidean)
and differential geometry have been used. The present paper suggests to employ
line geometry as a tool which is both interesting and applicable to a number
of problems in Computer Vision. Relations between vision and line geometry
are not entirely new. Recent research on generalized cameras involves sets of
projection rays which are more general than just bundles [1, 7, 18, 22]. A beautiful
exposition of the close connections of this research area with line geometry has
recently been given by T. Pajdla [17].
The present paper deals with the problem of understanding and reconstruct-
ing 3D shapes from 3D data. The data are assumed to be of a surface-like nature
— either a cloud of measurement points, or another 3D shape representation
such as a triangular mesh or a surface representation in parametric or implicit
form — and we assume that we are able to obtain a discrete number of points on
the surface and to estimate surface normals there. We are interested in classes of
surfaces with special properties: planar, spherical and cylindrical surfaces, sur-
faces of revolution and helical surfaces; and the more general surface classes of
canal, pipe, and developable surfaces. For applications in CAD/CAM it is essen-
tial that such special shapes are not represented by freeform surfaces without
regard to their special properties, but treated in a way more appropriate to their
‘simple’ nature.
Line geometry enters the problem of object reconstruction from point clouds
via the surface normals estimated at the data points. In fact, modern 3D pho-
tography and the corresponding software delivers such normals together with
2 Helmut Pottmann et al.

the data points. It turns out that the surface classes mentioned above can be
characterized in terms of their surface normals in an elegant, simple and compu-
tationally efficient way. Appropriate coordinates for lines (which yield a model
of line space as a certain 4-dimensional manifold embedded in R6 ) will allow to
classify point clouds and their normals (i.e., the so-called ‘normal congruence’)
by means of tools such as principal component analysis.
Previous work There is a vast body of literature on surface reconstruction.
Since we are interested in the reconstruction of special surfaces, we do not review
the part of literature which deals with the reconstruction of triangular meshes
or general freeform representations. In Computer Vision, recognition and recon-
struction of special shapes is often performed by methods related to the Hough
transform. Originally designed for the detection of straight lines in 2D, it re-
ceived much attention and has been generalized so as to be able to detect and
reconstruct many other shapes (see, e.g., [11, 14]). Pure Hough transform meth-
ods work in ‘spaces of shapes’ and quickly lead to high dimensions and reduced
efficiency. In order to avoid these problems, such tools are sometimes augmented
by methods from constructive geometry. This approach is already close to tech-
niques invented by the CAD community, which use geometric characterizations
of surfaces (e.g. by means of the Gaussian image) for data segmentation and
the extraction of special shapes (see the survey [24]). Many papers deal with
axis estimation of rotational surfaces, like [25]. See [8] for an overview on the
Hough transform, the RANSAC principle, and the least squares approach. In
the present paper, however, rotational surfaces occur only as a special case.
The use of line geometry for surface reconstruction has been introduced by
[20]. There cylinders, surfaces of revolution and helical surfaces are recognized
by the fact that their surface normals are contained in a so-called linear line
complex. In particular surfaces which can be moved within themselves in more
than one way (right circular cylinders, spheres, planes) are detected. The tech-
nique is extendable to surfaces which may be locally well approximated by the
surface types mentioned above [2, 13, 21].
Contributions of the present paper Inspired by the line geometric work on
reverse engineering of special shapes, our paper presents a broader line geometric
framework for the solution of problems in 3D shape understanding, segmentation
and reconstruction:
• We discuss a point model for line space, namely a certain 4-dimensional al-
gebraic manifold M 4 of order 4 in R6 . It is better suited for line geometric
approximation problems than the classical Klein model or the model used in
[20], which is limited to linear line complexes.
• This point model makes it possible to perform the basic shape recognition
tasks via principal component analysis (PCA) of a point cloud (contained in
M 4 ), which represents the estimated surface normals of the input shape. This
procedure is further improved here and unlike [20] is stable in all special cases.
• The idea of looking for surface normals which intersect makes it possible to
apply line-geometric methods to the problem of recognition and reconstruction
of canal surfaces (which are the envelope of a one-parameter family of spheres),
 Line Geometry for 3D Shape Understanding and Reconstruction 3

and of moulding surfaces. The latter denotes a certain class of sweep surfaces
which contains the developable surfaces and the pipe surfaces.
• The segmentation of composite surfaces into their ‘simple’ parts is addressed
here in so far as recognition of surface type is essential for segmentation. Our
algorithms may be included as a ‘black box’ into a segmentation algorithm.

1 The 4-dimensional manifold of lines in R6

This paragraph discusses a computationally attractive point model of (i.e., co-
ordinates in) the 4-dimensional manifold of straight lines in space. It is closely
related to the classical Klein quadric (i.e., Plücker coordinates). We think of an
oriented line L as one equipped with a unit vector l indicating the direction of
the line — so that there are two oriented lines for each line of space. Then L is
determined by l and the moment vector l̄, which is computed by means of an
arbitrary point x on L as l̄ = x × l. l̄ is independent of the choice of x on L. The
six numbers (l, l̄) are called normalized Plücker coordinates of L. ‘Normalized’
means that klk = 1; further, they satisfy the orthogonality condition l · l̄ = 0.
Conversely, any two vectors l, l̄ ∈ R3 which satisfy these two conditions deter-
mine a unique oriented straight line L in R3 , which has (l, l̄) as its normalized
Plücker coordinates.
If we do not distinguish between the two opposite orientations of the same
line, we may use all multiples of the pair (l, l̄) as coordinates of a line. Of course,
we still have the condition l · l̄ = 0. Such homogeneous coordinate vectors of lines
represent those points of five-dimensional projective space which are contained
in the Klein quadric M24 given by the equation (x, x̄) ∈ M24 ⇔ x · x̄ = 0. This
interpretation of lines is well studied in classical geometry, see [21].
The present paper pursues the following approach, which is closely related to
the Klein quadric. We use only normalized coordinate vectors, and so we identify
an oriented line with the point (l, l̄) in six-dimensional Euclidean space R6 . In
this way we obtain a mapping α of oriented lines to points of a 4-dimensional
manifold M 4 ⊂ R6 . M 4 is algebraic of degree 4, and is the intersection of the
cylinder Z 5 and the cone Γ 5 defined by
Z 5 : x2 = 1, Γ 5 : x · x̄ = 0.
We use the Euclidean distance of points in R6 in order to measure distances
between oriented lines G, H: If Gα = (g, ḡ) and Hα = (h, h̄), then
d(G, H)2 = (g − h)2 + (ḡ − h̄)2 . (1)
The reasons why we prefer this distance function are the following: On the one
hand, it is quadratic and thus lends itself to minimization. On the other hand,
in a neighbourhood of the origin of the coordinate system, (1) models a distance
of lines which is in accordance with visualization. The slight drawback that this
is no longer the case in regions far away from the origin is in fact not important
in most applications, as such applications very often define a natural region of
interest, where we can put the origin of our coordinate system.
4 Helmut Pottmann et al.

Remark 1. Another method for introducing coordinates and measuring distances

for lines, which has been used in the past, is to fix two parallel planes and
describe a line by the two intersection points with that plane (cf. [21]). This
leads to simpler formulae and a region of interest which is bounded by the two
fixed planes (in our case, this region is a sphere centered in the origin).

Classification of surfaces by normal congruences

The set of normals of a surface is called its normal congruence. Some surface
types are easily recognized from properties of their normal congruence: The
normals of a sphere pass through its center (they constitute a bundle with a finite
vertex), and the normals of a plane are parallel (they constitute a parallel bundle,
with vertex at infinity). These are the simplest examples; there are however
other interesting and practically important classes of surfaces which are nicely
characterized by their normal congruence. These classical results are basic to
shape understanding and reconstruction algorithms and thus we summarize them
in this section.
A uniform motion in 3-space, composed of a uniform rotation of unit angular
velocity about an axis and a translation of constant speed p along this axis is
called a helical motion of pitch p. If we choose a Cartesian coordinate system
with the x3 -axis being the axis of rotation, then the point (x1 , x2 , x3 ) will move
according to

x1 (t) = x1 cos t − x2 sin t, x2 (t) = x1 sin t + x2 cos t, x3 (t) = x3 + pt. (2)

In the case p = 0 we have a uniform rotation. For p → ∞ we get, in the

limit, a uniform translation. A surface swept by a curve under a helical motion
is called helical surface. As special and limit cases for p = 0 and p = ∞ we
get surfaces of revolution and cylinder surfaces, respectively. We say that these
surfaces are kinematically generated. However if we speak of a helical surface we
always mean part of a complete helical surface as defined above, and analogously
for other adjectives, like rotational/cylindrical/spherical/planar. Closely related
to helical motions are linear complexes, which are certain three-parameter sets
of lines defined by linear equations, and which are discussed in more detail in
[21]. A line L with Plücker coordinates (l, l̄) is contained in the complex C with
coordinates (c, c̄) if and only if

L ∈ C ⇐⇒ c̄ · l + c · l̄ = 0. (3)

Obviously the lines of the complex C defined by (3) correspond to those points
of R6 which are both contained in M 4 and fulfill (3), i.e., they lie in M 4 and in
the hyperplane H 5 : c̄ · x + c· x̄ = 0. Note that H 5 passes through the origin. The
set Cα is a certain 3-dimensional manifold. A linear complex C is called singular
if c · c̄ = 0 (then there is a line A with Aα = (c, c̄), and C consists of all lines
which intersect A).
 Line Geometry for 3D Shape Understanding and Reconstruction 5

The sets of lines which correspond to the intersections of M 4 with other d-

dimensional subspaces H d also play important roles and have special names.1
These line sets are employed by the following classification result (see e.g. [21])

Proposition 1. The normals of a given surface are contained in a linear com-

plex C = (c, c̄) if and only if the surface is helical or rotational or cylindrical.
The complex C is regular for a helical surface and singular otherwise. The axis
of C is at infinity (i.e., c = 0) for cylindrical surfaces. Parts of surfaces which
are kinematically generated in more than one way are characterized as follows:
The normals of a right circular cylinder are contained in a linear congruence;
the normals of spherical and planar surfaces are contained in a bundle.

A list of surface classes and their normals Surfaces of revolution (Fig.

1f) are envelopes of a one-parameter family of spheres, the centers of which lie
on the axis of rotation. More generally the envelope of a smooth one-parameter
family of spheres is called a canal surface (see Fig. 1a). The midpoints of these
spheres form the surface’s spine curve. If a sphere touches the envelope surface
it does so along a circle.
If the spheres are of constant radius, one obtains a pipe surface (see Fig.
1c). Pipe and canal surfaces appear in CAD for example as blending surfaces.
Obviously pipe surfaces play an important role in places like oil platforms and
refineries. As-built reconstructions of these shapes from 3D data have a number
of applications and thus this topic already received some attention (see e.g.
[16]). Viewpoint-invariant recognition of canal surfaces from images has been
addressed e.g. by Pillow et al. [19].
The envelope of a sphere under rotation about an axis (not through the
sphere’s center) is a torus (Fig. 1g), which may be generated as a canal surface
also in another way. In general, surfaces, which are canal surfaces in two ways,
are called Dupin cyclides (see Fig. 1e). They are well known algebraic surfaces
of degree 3 or 4, and may be described as images of tori under inversion. Their
use in geometric modelling is described e.g. in a survey article by W. Degen [4].
Pipe surfaces are also traced out by a circle p which moves such that its
center runs along the spine curve s(t), and its plane U (t) is always orthogonal
to that curve. Rotation about the spine curve has no influence on the resulting
surface, but we would like to specify the movement in a more precise way: We
assume that a coordinate frame attached to the circle moves such that its angular
velocity is minimal (i.e., no unnecessary rotations about the spine curve occur;
see e.g. [12]). This is the case of the rotation-minimizing frame, where trajectories
1
For d = 4 we get linear line congruences (e.g. a hyperbolic linear congruence consists
of all lines which meet two given lines). The case d = 3 consists of three subcases:
(i) the bundle of lines incident with a given point; (ii) the field of lines contained in
a given plane; and (iii) a regulus, which is one of the two one-parameter families of
lines contained in a ruled quadric. In the cases (i) and (ii) H 3 is contained in Γ 5 ;
whereas in case (iii) the set Γ 5 ∩ H 3 is a quadratic cone. For d = 2 it may happen
that H 2 is contained in Γ 5 , in which case the corresponding line set is a pencil,
consisting of lines in a fixed plane U which pass through a point p ∈ U .
6 Helmut Pottmann et al.

(a) (b) (c) (d)

(e) (f) (g) (h)

Fig. 1. (a) Canal surface, (b) Moulding surface, (c) Pipe surface, (d) Developable
surface, (e) Dupin cyclide, (f) Surface of revolution, (g) Torus, (h) Cone of revolution.
Both the surface and the set I(x) are shown.

of points in the plane U (t) are orthogonal to U (t). A surface generated by any
profile curve p (not necessarily a circle) during this motion is called a moulding
surface (see Fig. 1b). It has the property that all positions p(t) of the profile
curve are principal curvature lines.
If the profile curve p is a straight line, the moulding surface is developable and
can be mapped isometrically into the Euclidean plane. All points of a generator or
ruling p(t) have the same tangent plane. Thus, such a surface is also the envelope
of a one-parameter family of planes. Special cases are cones and cylinders, but
in general the lines p(t) are tangent to some space curve (the line of regression),
which is a singular curve on the surface (see Fig. 1d).
Let us now describe how to characterize these surfaces via their normal con-
gruences. It is well known that the normal congruence of any surface Φ, which
is not planar or spherical, can be decomposed into two families of developable
surfaces. These are formed by the surface normals along the principal curvature
lines. For canal surfaces one of these families consists of cones (or possibly cylin-
ders) of revolution, and the vertices of these cones lie on the spine curve (see Fig.
1a). For pipe surfaces, these cones have an opening angle of 180 degrees (i.e.,
they are line pencils; see Fig. 1c). For developable surfaces the cones become
cylinders of infinite radius (i.e., they are pencils of parallel lines; see Fig. 1d).
In order to exploit these facts computationally (cf. Sec. 3), we define the
set I(x) of locally intersecting normals at a point x of a surface Φ: We pick
a neighborhood N (x) ⊂ Φ of x and look for points y in N (x) whose normals
intersect the normal at x. These normals form the set I(x). Interesting are those
cases where we can identify two specially shaped components of I(x).

Proposition 2. The following table enumerates special surface classes and the
shapes of the two components of I(x) for any surface point x (cf. Fig. 1).
 Line Geometry for 3D Shape Understanding and Reconstruction 7

cone planar pencil parallel

arbitrary canal surface moulding surface pipe surface developable surface
cone cyclide rotational surface torus rotational cone
Conversely, if for all x we can identify two components of I(x) as listed, the
surface is of the associated type.
Here, ‘arbitrary’ means that no condition is imposed on this component, ‘parallel’
means that the lines of the component are parallel, and ‘planar’ means that
the lines are contained in some plane. The table can be used as follows. If,
for example, we find that I(x) contains, for all x, a parallel component, and a
conical component, the surface is a cone of revolution (as a side-effect the conical
component itself must be part of a cone of revolution; see Fig. 1h). Note that the
second row of the table, except for the cyclides, contains only rotational surfaces
already addressed by Prop. 1. This is the reason why adding new rows in the
table above does not lead to anything new.

2 Principal component analysis (PCA) on the surface

normals for surface recognition and reconstruction

Basic algorithm and PCA We would like to approximate a set of lines Li ,

i = 1, . . . , N by a linear line complex C ∗ with coordinates (c∗ , c̄∗ ). It will be
sufficient to consider one orientation for each line, so we compute normalized
Plücker coordinates Li α = (li , l̄i ) ∈ R6 .
A first approximation method is to compute the unique hyperplane H 5
through the origin which minimizes the squared sum of distances from the points
Li α. With H(x, x̄) = c̄ · x + c · x̄, the Euclidean distance of a point (p, p̄) from
the hyperplane H 5 : H(x, x̄) = 0 simply equals H(p, p̄), if
c2 + c̄2 = 1. (4)
N
Thus we find H 5 by minimizing F (c, c̄) = i=1 [c̄ · li + c · l̄i ]2 under the con-
P
straint (4). We consider both c and c̄ as column vectors and write F in the form
hci
F (c, c̄) = [cT c̄T ] · A · , with a certain symmetric 6 × 6-matrix A. It follows
c̄
easily from the Lagrangian multiplier rule that the minimizer (c∗ , c̄∗ ) is given
by an eigenvector of A, which belongs to A’s smallest eigenvalue λ∗ , and which
is normalized according to (4) (in this case we also have F (c∗ , c̄∗ ) = λ∗ ).
According to Prop. 1, to a linear complex C belongs a helical motion. Its
pitch p and the Plücker coordinates (a, ā) of its axis can be computed by
p = (c∗ · c̄∗ )/(c∗ )2 , (a, ā) = (c∗ , c̄∗ − p c) (cf. [21]). A small pitch p indi-
cates that the lines Li belong to a surface of revolution. We may then repeat the
approximation process with the additional side condition that c · c̄ = 0 (accord-
ing to Prop. 1). Large values of p indicate that the lines Li belong to a cylinder.
This circumstance is also detected by the fact that all Li are orthogonal to a
fixed line. If we decide we want the data to be approximated by a cylinder, we
may repeat the minimization with the additional side condition that c = 0.
8 Helmut Pottmann et al.

So far we have shown how to find the motion which generates the surface.
For the actual computation of a profile curve we use methods available in the
literature [20, 21, 23, 24].
Two small eigenvalues of A indicate that the points Li α ∈ R6 are almost
contained in a subspace of dimension 4. The normals Li are then almost con-
tained in a linear congruence. By Prop. 1 this means that the normals belong
to a right circular cylinder. As a side-effect, all of them must now intersect the
cylinder’s axis orthogonally.
Three small eigenvalues of A imply that the points Li α are almost contained
in a 3-dimensional subspace, which indicates a spherical or planar surface. Note
however that planes are also detected by a Gaussian image [24] or a Hough
transform for planes in 3-space.
Remark 2. The procedure above is a principal component analysis of the set of
image points Li α ∈ R6 . Without going into details we remark that it may be
refined in order to be more robust e.g. by incorporating a weighted least squares
approximation which iteratively downweights outliers, or the RANSAC principle
(see [5] and references therein).
Refined algorithm So far we computed an approximating complex C by
finding a hyperplane H 5 (carrying Cα) such that the sum of squares of distances
of the points Li α from H 5 is minimized. This is not the same as minimizing
distances of the points Li α from the set Cα = M 4 ∩ H 5 . Actually we should have
minimized the geodesic distance of the points Li α from the set Cα = H 5 ∩ M 4
within the manifold M 4 . We will take this into account by considering the angle
of intersection between H 5 and M 4 . For a point (l, l̄) = Lα in M 4 consider M 4 ’s
tangent 4-space T 4 and let φ = ∠(T 4 , H). dH denotes the distance of Lα from
H 5 . Then the value dC = dH / sin φ is an estimate for the geodesic distance of
Lα from Cα = M 4 ∩ H 5 . There is the following result:

Lemma 1. The angle φ used in the definition of dC is given by

n n 2
cos2 φ = (c̄ · l)2 + (c̄ · l̄ + c · ln )2 , with (ln , l̄ ) = 1/(l2 + l̄ ) · (l, l̄).

Proof. We note that the 2-dimensional normal space to M 4 = Z 5 ∩ Γ 5 in the

point Lα = (l, l̄) is spanned by the normal vectors a to Z 5 and b to Γ 5 . These
are given by a = (l, o) and b = (l̄, l), respectively. Note that a · b = 0, and that
kak = 1, whereas b is not yet a unit vector in R6 . In order to achieve this, we
n n
define ln and l̄ as above, and replace b by bn = (l̄ , ln ). Now (c, c̄) is a normal
vector of H 5 and satisfies (4). It follows that cos2 φ = (a · n)2 + (bn · n)2 , and
the result follows by expanding the definitions. t
u

Minimizing the distances dC instead of orthogonal distances dH means min-

imizing the function N
X (c̄ · li + c · l̄i )2
F1 (c, c̄) = n
i=1
1 − (c̄ · li )2 − (c̄ · l̄i + c · lni )2
under the side-condition (4). This is a nonlinear optimization problem which
we solve via a weight iteration: We initialize the algorithm with the minimizer
of F . Then, using the solution (c∗ , c̄∗ ) from the previous step, we minimize
 Line Geometry for 3D Shape Understanding and Reconstruction 9

(a) (b)

Fig. 2. Reconstruction of axis and meridian curve of a pottery object. Left: Axes re-
constructed from parts (1,2) and from entire object (3). Center: Reconstructed merid-
ian curve. Right: Colour coded deviations of original from reconstructed ideal surface
(maximum deviation: 2.6% of diameter).

Fig. 3. Left: Reconstruction of axis and two helical paths for helical gears. Right: Photo
of helical gears.

n
F2 (c, c̄) = N 2 ∗ 2 ∗ ∗ n 2
P
i=1 wi (c̄ · li + c · l̄i ) , with 1/wi = 1 − (c̄ · li ) − (c̄ · l̄i + c · li ) .
This procedure is motivated by the fact that different angles φi at points Li α
give different weights to distances dH,i in the minimization of F . The unbalanced
situation is corrected via the weight iteration described above. The iteration
terminates if the change in F1 in one iteration step is below some given threshold.

3 Examples

We used a Minolta VI-900 3D laser scanner to obtain point data and tested the
effectiveness of our algorithms on them. Point clouds were thinned if necessary
and triangulated. Surface normal vectors have been estimated by local regression
planes. In Figures 2, 3, and 4, data points have been rendered as small balls.
Surface type by PCA: Rotational surfaces Fig. 2 shows the procedure
of reconstructing the axis and meridian curve of a near-rotational surface. Data
have been obtained by scanning the outer surface of a late Hallstatt pottery
object manufactured in approx. 550 B.C. without the use of a pottery wheel.
The procedure of Sec. 2 has been applied to estimate a linear line complex which
10 Helmut Pottmann et al.

fits the surface normals best. With d as the diameter of the point cloud, we found
|p| ¿ d, so the data points in question may be approximated by a surface of
revolution. The meridian curve (Fig. 2, center) was found by rotating the point
cloud into a plane through the axis of rotation (a) and approximating (a) by a
smooth curve (b). The deviations of the original cloud from the reconstructed
ideal surface of revolution are shown by Fig. 2, right. Figure 2, left, shows how
the estimated axis (3) changes if we use only parts of the data available (dotted
lines 1,2). As is to be expected, accuracy decreases if we use small meridian
strips.
Surface type by PCA: Helical surfaces Fig. 3, left shows a scan of four teeth
of a helical gear wheel as an example of a surface which is known to be helical.
The point data have been obtained in one pass of scanning. The underlying
helical motion, defined by axis and parameter, has been reconstructed in the
expected way.
Surface type by PCA: Freeform surfaces The human body does not
possess mathematically exact helical surfaces. However, we studied the following
interesting example: Fig. 4, left, shows a scan of a trochlea tali, i.e., the distal
interface of the ankle joint. The closest helical surface computed by the algorithm
of Sec. 2 is not a surface of revolution. This piece of information is important
when studying the relative motion of the talus (ankle bone) with respect to the
tibiofibular (lower leg) system. We might ask whether the trochlea tali is close
enough to a mathematical helical surface to be called helical. This turns out to
be not the case, as can be seen from computing the closest helical surfaces to
surface strips and comparing the results. The axes corresponding to four strips
together with the axis corresponding to the entire data set are indicated in Fig. 4.
Surface type by I(x) In order to recognize even more surface classes, we
consider the sets I(x) of locally intersecting normals. We compute them as fol-
lows: The surface normal spanned by the point x0 and the vector n0 is contained
in I(x), if and only if det(x0 − x, n, n0 ) = 0 (n being the normal vector at x).
The discrete version of this is that the normals at the endpoints of an edge of a
triangulation of our point cloud are in I(x), if the determinant mentioned above
changes its sign along that edge.
Our examples — scans of plaster and wood models — are shown by Fig. 4,
where the data points whose normals contribute to I(x) are indicated in black.
We can see that I(x) consists of two components, one of which may be an arc
(in case of a canal surface) or a straight line (in case of a developable surface).
In order to apply Prop. 2, we have to test whether one or more of the two
components of I(x) are conical or planar. If Li = (li , l̄i ) (i = 1, . . . , N ) are the
surface normals of such a component, we have to find either a point incident with
all of them or a plane which contains them. There are several ways to do this.
One would be to consider the lines Li as surface normals and to try to reconstruct
the corresponding surface type according to Sec. 2. This, however, would lead
to numerical difficulties due to thinness of data. Therefore we recognized conical
components with vertex s in the following way: Incidence of the point s with the
line (li , l̄i ) is characterized by l̄i = s × li . Thus finding a point s ‘as incident as
 Line Geometry for 3D Shape Understanding and Reconstruction 11

Fig. 4. Left: Computation of axes of nearest helical surfaces for strips 1–4 and also the
entire data set (shown in bold) of a trochlea tali. Right: The sets I(x) for, counting
clockwise from top, a surface of revolution, the developable oloid, and for the plaster
model of the Dupin cyclide shown in [10], Fig. 229a.

possible’ with the lines Li means minimizing

X
G(x) = wi (l̄i − s × li )2 , (wi ≥ 0) (5)

with weights wi which determine the influence of the single lines. Minimizing
(5) is standard, and iteratively downweighting
√ outliers makes this method sta-
ble. Comparison of the minimal value of G with the size of the point cloud
determines conicality. Planar components of I(x) are detected in a similar way.
Note that line pencils are both planar and conical.
The sets I(x) of Fig. 4 have, counting clockwise from the top, 1, 0, and
2 (non-planar) conical components, whereas the number of planar components
equals 1, 1, and 0, respectively. It follows that these surfaces are a rotational
surface, a moulding surface, and a Dupin cyclide. Additional information on the
parallelity of surface normals along one component of I(x) shows that the right
hand surface shown by Fig. 4 is a developable one.
Conclusion and Future Research We have shown how techniques from
classical line geometry can serve for recognizing and reconstructing special sur-
faces. Other applications, such as ruled surface approximation and the computa-
tion of approximating line congruences can benefit from the use of the embedding
of line space into R6 . Active B-spline curves and surfaces can be used efficiently.
This is a consequence of the fact that finding footpoints on M 4 corresponds to
finding the zeros of a fourth order polynomial, and is in fact equivalent to the
footpoint problem for planar ellipses.
Future research will also address how to use the bisecting linear complex (cf.
[21], p. 166) for checking and improving point correspondences in 3D registration
problems, especially when looking for a good initial position.
Acknowledgements This work was supported by the Austrian Science Fund
(FWF) under grants P16002 and P15911; and the innovative project ‘3D tech-
nology’ of Vienna University of Technology. We want to express our thanks to
Landesarchäologie Salzburg for the archaeological pottery artifact.
12 Helmut Pottmann et al.

References
1. Benosman R., Kang S. B. (editors): Panoramic Vision: Sensors, Theory and Ap-
plications, Springer Verlag, 2001.
2. Chen H.-Y., Lee I.-K., Leopoldseder S., Pottmann H., Randrup T., Wallner J.:,
On surface approximation using developable surfaces, Graphical Models and Image
Processing, 61:110–124, 1999.
3. Cipolla R., Giblin P.: Visual Motion of Curves and Surfaces, Cambridge UP, 2000.
4. Degen W.: Cyclides, In Handbook of Computer Aided Geometric Design, G. Farin,
J. Hoschek and M.-S. Kim, eds., Elsevier, pp. 575–601, 2002.
5. De la Torre, F., Black M. J.: Robust principal component analysis for Computer
Vision,Proc. 8th Int. Conf. on Computer Vision, pp. 362–369, 2001.
6. Faugeras O.: Three-dimensional Computer Vision: A Geometric Viewpoint, MIT
Press, Cambridge, MA, 1993.
7. Gupta R., Hartley R.: Linear pushbroom cameras, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 19(9):963–975, 1997.
8. Halı́ř R.: Estimation of the axis of rotation of fragments of archaeological pottery.
Proc. 21st Workshop Austrian Assoc. for Pattern Recognition, Hallstatt 1997.
9. Hartley R., Zisserman A.: Multiple View Geometry in Computer Vision, Cambridge
Univ. Press, Cambridge, UK, 2000.
10. Hilbert D., Cohn-Vossen S., Anschauliche Geometrie. Springer, 1932. Reprinted
1996. Translated as: Geometry and the Imagination, American Math. Soc. 1999.
11. Illingworth J., Kittler J.: A survey of the Hough transform, Computer Vision,
Graphics and Image Processing, 44:87–116, 1988.
12. Jüttler B., Wagner M.: Kinematics and animation, In Handbook of Computer Aided
Geometric Design, G. Farin et al. eds., Elsevier, pp. 723–748, 2002.
13. Kós G., Martin R., Várady T.: Recovery of blend surfaces in reverse engineering,
Computer Aided Geometric Design, 17:127–160, 2000.
14. Leavers V.: Which Hough transform? CVGIP:Image Understanding 58:250-264,1993.
15. Mundy J., Zissermann A.:Geometric Invariance in Computer Vision, MIT Press, ’92.
16. Navab N.: Canonical representation and three view geometry of cylinders, Intl.
Archives of the Photogrammetry, Remote Sensing and Spatial Information Sci-
ences, Commission III, Vol. XXXIV, Part 3A, pp. 218–224, 2002.
17. Pajdla T.: Stereo geometry of non-central cameras, PhD thesis, CVUT, 2002.
18. Peleg S., Ben-Ezra M., Pritch Y.: Omnistereo: panoramic stereo imaging, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 23(3):279-290, 2001.
19. Pillow N., Utcke S., Zisserman A.: Viewpoint-invariant representation of general-
ized cylinders using the symmetry set, Image Vision Comput., 13:355–365, 1995.
20. Pottmann H., Randrup T., Rotational and helical surface reconstruction for reverse
engineering, Computing, 60:307–322, 1998.
21. Pottmann H., Wallner J.: Computational Line Geometry, Springer 2001.
22. Seitz S. M.: The space of all stereo images, Proceedings ICCV 2001, pp. 26–33.
23. Várady T., Benkő T., G. Kós T.: Reverse engineering regular objects: simple seg-
mentation and surface fitting procedures, Int. J. Shape Modeling, 4:127–141, 1998.
24. Várady T., Martin R.: Reverse Engineering, In Handbook of Computer Aided Ge-
ometric Design, G. Farin et al., eds., Elsevier, pp. 651–681, 2002.
25. Willis A., Orriols X., Cooper D.: Accurately Estimating Sherd 3D Surface Geom-
etry with Application to Pot Reconstruction. CVPR Workshop, 2003.