Professional Documents
Culture Documents
Correspondence: Computation Surface Geometry and Segmentation Using Covariance Techniques
Correspondence: Computation Surface Geometry and Segmentation Using Covariance Techniques
IEEE TRANSACTIONS ON PAITERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 16, NO. I 1, NOVEMBER 1994
Correspondence
Computation of Surface Geometry and
Segmentation Using Covariance Techniques
Jens Berkmann and Terry Caelli
Abstract-In this correspondence, the application of covariance techniques to surface representationof 3-D objects is discussed and such ways
of computing surface geometry are compared with traditional methods
using Differential Geometry. It is shown how the covariance method
provides surface descriptors that are invariant to rigid motions without
explicitly using surface parameterizations or derivatives. Analogous covariance operators for both the Gauss and Weingarten maps are defined
and a range image segmentation technique is presented that labels pixels
as jump or crease discontinuities or planar, parabolic or curved region
types.
Index Terms- Differential Geometry, covariance, Gauss map, image
segmentation, jump map, crease map, region segmentation.
I. INTRODUCTION
Object recognition systems based upon range data typically compute surface shape interms of surface normals or curvatures as
developed from Differential Geometry (see, for example [l], [2],
[3], [5]). However, problems do occur with implementing these descriptors, notably, the need for surface parameterization and adequate
smoothness to support the use of Differential operators. Further to
this, such curvature measures are not necessarily robust under additive
noise as they involve nonlinear combinations of partial derivatives.
For these reasons we have further explored an altemate signal
processing basis for computing local shape measures4ovariance (as
recently used by Liang and Todhunter [4])--and show how invariant
surface shape operators can be developed with this statistic.
The covariance approach dispenses with surface parameterization
and provides invariant descriptions of shape via the eigenvalues of covariance matrices of different orders. The aim of this correspondence
is to further investigate the properties of the covariance approach
and to propose definitions of the Gauss and the Weingarten maps
derived from covariance matrix operations. Finally, we propose a new
range image segmentation technique using this covariance approach
to segment the surface into planar, parabolic, and curved regions as
well as jump and crease discontinuities.
--2
. (E* - %
)proj,
I
st
IIC = J
APPROACH
11. COVARIANCE
Following Liang and Todhunter [4] we define the local (first order)
surface covariance matrix ( C I )as:
(3)
.CII .c
(5)
as a second fundamental form (Weingarten map) based on covariance methods with the defined covariance matrix according to
(2) and a chosen unit vector in the tangent plane. Analogous to
classical computations of surface geometry, we define the principal
directions as those directions that are given by the eigenvectors of
this covariance matrix. In contrast, Liang and Todhunter [4] estimated
principal directions by calculating the covariance matrix of a given
normal map. They selected the two eigenvectors of the covariance
matrix corresponding to the two largest eigenvalues as estimations
of principal directions.
An altemate way of computing such principle directions, since they
lie on the tangent plane anyway, is to calculate the covariance matrix
from the projections of the normal vectors, within a neighbourhood of
a point, onto the tangent plane and define the principal directions as
--
1115
TRANSACTIONS ON PAITERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 16, NO. 11, NOVEMBER 1994
the eigenvectors of this two-dimensional covariance matrix. In addition, this definition enables us to establish a segmentation technique
for planar, parabolic, and curved regions based on the observation
of the characteristic clustering of the projected normals onto the
tangent plane.
Consequently, we have an analogous operator to the Gauss map
of classical differential geometry. In this way we define the twodimensional covariance matrix:
creases. The projection onto the tangent plane plays a more important
role in the following stage of region segmentation.
1116
IEEE TRANSACTIONS ON PATIERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 16, NO. 11, NOVEMBER 1994
method and image, which gave best results in terms of the entropy
measures before comparing them.
IV. DISCUSSION
Fig. 1. Firsr row: Original range image. Second row: Combined jump-and
crease-edge maps for SDG (left) and covariance (right) methods. Jumps
have been assigned the brighter color.Third row: Final segmentation for SDG
(left) and Covariance (right) methods. Planar, parabolic, curved regions and
discontinuities are shown, where the brightness increases in the above order
(planar corresponding to black). For reasons of better visibility the jump and
crease labels were joined (white).
exhibited, even without any pre- or postprocessing, quite acceptable
regionness and smoothness results. The thresholds for both methods
were chosen by observing the results from a wide range of thresholds
and selecting results that optimized, visually, the fit of surface parts
to the known [5] surface geometry types: jumps, creases, planar,
parabolic and curved. However, to provide an objective measure of
the achieved degree of regionness, we computed an average entropy
(uncertainty) measure as follows. A 7 x 7 window was moved over
the full image and, at every pixel, the relative frequencies of labels [5]
were computed. From these values probabilities and the associated
information statistic H = - p, log p, was computed. The average
of this measure constituted the average entropy measure: Ha,,. We
found that for our examples the average entropy Ha,,, of the reference
method (final segmentation) was in between 1.4 and 2.5 times larger
than the average entropy of the covariance approach (see Fig. I).
A final note should be made on the choice of the thresholds.
Automatic determination of thresholds, in particular, the thresholds
for the discontinuity detection can be accomplished by using, for
example, histogram information or even expert knowledge-though
this has not become, as yet, a standard procedure in the literature. For
this reason and for reasons of a fair comparison of our method with
already established methods, we selected threshold values, for each
ACKNOWLEDGMENT
The authors wish to thank S. Berkmann for helping with the
simulations.
REFERENCES
P. J. Besl and R. C. Jain, Invariantsurface characteristicsfor 3D object
recognition in range hages. Computer Vision, Graphics and Image
Processing, vol. 33, i o . 1, pp. 33-80, 1986.
Ting-Jun Fan, Gerard Medioni, and Ramakant Nevatia, Recognizing
3-D objects using surface descriptions, IEEE Trans. Pattern Anal.
Machine Intell., vol. 11, no. 11, pp. 1140-1156, 1989.
Ani1 K. Jain and Richard Hoffman, Evidence-B;lsed Recognition of
3-D Objects, IEEE Trans. Pattern Anal. Machine Intell., vol. 10, no.
6, pp. 783-801, 1988.
P. Liang and J. S. Todhunter, Representation and recognition of
surface shapes in range images,Computer Vision, Graphics and Image
Processing, vol. 52, no. 10, pp. 78-109, 1990.
Naokazu Yokoya and Martin D. Levine, Range image segmentation
based on differentialgeometry: A hybrid approach,IEEE Trans. Pattern
Anal. Machine Intell., vol. 11, no. 6, pp. 643449, 1989.