CS273: Algorithms for Structure Handout # 8

and Motion in Biology

Stanford University Thursday, 22 April 2003

Lecture #8: 22 April 2004

Topics: Shape matching and structural comparison
Scribe: Nina Singhal

1 Introduction
In protein structure comparison, we are trying to determine how similar two structures
are. Here, we discuss four methods for measuring the similarity between two geometic
shapes.

2 Hausdorff Distance
The Hausdorff Distance is commonly used in computer vision. In that field, a typical
problem is that you are given an image and a model of what you want to match to. The
goal is to find all the locations in the image which match the model. This is similar to the
problem of matching protein motifs within protein sequences. This distance is different
from some of the previously discussed measures, because in this case, instead of forming
a one-to-one mapping between the two, we allow a many-to-many correspondence. Often
times, it is easier to build a many-to-many correspondence, since if we wish to change
one assignment, we no longer have a cascade of other assignments which now also need
to be redone.

2.1 Mathematical Definition

We are given two point sets A = {a1 , a2 , . . . , an } and B = {b1 , b2 , . . . , bm } in E 2 . The
one-sided Hausdorff distance from A to B is defined as:

δ̃H (A, B) = max min ka − bk (1)

a∈A b∈B

The bidirectional Hausdorff distance between A and B is then defined as:

δH (A, B) = max(δ̃H (A, B), δ̃H (B, A)) (2)

For fixed A and B, this can easily be computed in time O((n + m) log(n + m)) using
Voronoi diagrams. Sometimes, the one-sided distance is preferable, as in the case of
partial matching of models to images (under occlusions, etc.).
2 CS273: Handout # 8

Figure 1: A lower envelope surface

2.2 Variations
In practice, taking the maximum of all the distances is dangerous because possible outliers
in one set can then greatly impact the Hausdorff distance. We can compute the fractional
Hausdorff distance, in which say 90% of the points in A have that distance or less to some
point in B.
We can also allow one set of points to be moved by a group of transformations G, for
example translations or rotations. This is typically a much harder problem. The distance
is then defined as:
δ̃H,G = min max min ka − T (b)k (3)
T ∈G a∈A b∈B

2.3 Translation example

Suppose we wish to calculate the Hausdorff distance between A and some translation t of
the points in B. Consider the points in A to be points on a plane. The Voronoi surface
of A is a conical piecewise surface, where each cone is at a 45◦ angle to the plane. Since
the cones are at 45◦ , the vertical distance from any point up to the cone is the same as
the distance to the apex A. The minimum distance to any point in A can be defined by
the lower envelope surface of the cones, shown in figure 1. The distance from any point
x would then be
d(x) = min kx − ak . (4)
a∈A

If we consider a translation t of B, then

δb (t) = min ka − (b + t)k = min k(a − b) + tk = d−b (t). (5)

a∈A a∈A

The directed Hausdorff distance between A and some translation t of B is then defined
as
f (t) = δ̃H (B + t, A) = max δb (t) (6)
b∈B
CS273: Handout # 8 3

Figure 2: A 1-D example of calculating distance transforms

This is the same as taking the upper envelope of the m Voronoi surfaces, A − b1 , A −
b2 , . . . , A − bm . The running time of an algorithm based on this approach is O(nm(n +
m)polylog(n + m)). In general, this doesn’t work well unless there are relatively few
points. And, if we also allow rotations, or move the points to 3 − D, the computation
time, though still polynomial, becomes too expensive.
However, graphics cards can compute quantized approximations to these lower and
upper envelopes using a Z-buffer. These advances in hardware make these methods more
practical.

2.4 Raster Hausdorff

It is possible to compute distance transforms on a grid given an image. This transform
efficiently computes how far each grid point is from the given points in the set. For
example, as in figure 2 in 1-D, we can compute this grid in two passes using fast marching
or level sets. We begin by initializing all the points in our set to 0 and all other points in
the grid to ∞. Then, we do a pass to the right, where we maintain a counter, initially set
to ∞. Every step right, we increment the counter, and if we encounter a zero, we set the
counter to zero. If the value in the grid is greater than the counter, we set the grid value
to the counter value. We follow with a similar pass to the left. In this same manner, we
can compute the distance transform for a 2-D set of points by doing 4 passes, %, &, -,
., and in 3-D with 8 passes.

2.5 Fast Hausdorff Search

In practice, it is common to use a branch and bound hierarchical search of the transfor-
mation space. If we consider 2-D transformation space of translation in x and y, then
the rate at which the Hausdorff distance can change is linear with the translation. So,
we can do a quad-tree decomposition, where we compute the distance for the transform
at the center of each cell. If the distance minus the cell half-width is larger than our
current best estimate, then we can rule out that cell; otherwise, we subdivide the cell
and consider the children, as in figure 3. A guaranteed, or admissable, search heuristic
bounds how good the answer could be in the unexplored region. These search heuristics
can’t miss any answers, but in the worst case, won’t rule out any of the search space. In
practice, however, we can rule out the vast majority of transformations. In fact, we can
4 CS273: Handout # 8

Figure 3: Branch and bound techniques for fast Hausdorff search

use even simpler tests than computing the distance at each cell center.

2.6 Reference Points

For some shapes we can define reference points where if you align these points, the
rest of the shapes will align reasonably well. These schemes can give constant factor
approximations to the Hausdorff distance.
For example, for translation in 2 − D, the lower left corner of the bounding box of the
shape serves as a good reference point. Call δ = δH (A, B) the true Hausdorff distance
between two shapes. Also, as in Figure 4, label the lower left hand corner of the bounding
boxes of shape A and shape B by rA and rB respectively. Then, the maximum x distance
between rA and rB is bounded by δ, since each point in B is to the right of rB but each
point in A is at most δ away from some point in B, including the point at rA . The same
holds for the maximum y distance between rA and rB .
Thus, we can conclude that
√
krA − rB k ≤ 2δ. (7)

Thus, the Hausdorff distance between A and B 0 , where B 0 is where we match the lower
CS273: Handout # 8 5

Figure 4: Aligning shapes by their reference points

left bounding box corners is

√
δH (A, B 0 ) ≤ δH (A, B) + δH (B, B 0 ). ≤ ( 2 + 1)δ (8)
with the first inequality holding because the Hausdorff distance satisfies the triangle
inequality. This shows that aligning the bottom left corner of the bounding boxes results
in a constant-factor approximation to the true Hausdorff distance. We should also be able
to improve on this alignment by local resampling, since we know the optimal alignment
is close by.

3 Fréchet distance
One downside of the Hausdorff distance is that it may call things similar which don’t
seem alike. For example, in figure 5, the two shapes are not alike, but since any point
in one is very close to some point in the other, the Hausdorff distance will be small.
We may wish for the correspondence between the two shapes to reflect the underlying
connectivity of their shapes. The Fréchet distance takes this into account. Imagine a
man walking his dog, and the two paths taken by each. We assume that the man and
his dog may only move forward on their paths. The goal is to find a mapping through
time of the two paths, such that the maximum distance between them is minimum. In
other words, we try to find the minimum length leash necessary between the man and
the dog. The equation for the Fréchet distance is:
δF (f, g) = inf max kf (α(t)) − g(β(t))k (9)
α,β t∈[0,1]
6 CS273: Handout # 8

Figure 5: Two different curves

Figure 6: Fréchet decision problem

where f and g are the two shapes and α and β are the two parameterizations.
However, the problem of trying to find a correspondence between the two paths is
hard. We can instead switch from the optimatization problem of finding the minimum
length of the leash to a decision problem which asks whether a parameterization exists
for a leash of a certain length, x. This can be solved by graphing the two paths and
coloring in white any points which are within x of each other, as in figure 6. α and β
can be any x- and y- monotone paths which connect (0, 0) and (n, m) and always stay in
the white regions. Using this decision procedure, we can find the minimum length leash
using binary search. The total running time for this algorithm is O(mn log(mn)).
CS273: Handout # 8 7

Figure 7: Morphing one shape into another

Figure 8: The Earth Mover’s Distance

4 Morphing distance
The morphing distance is a measure which computes the cost of changing one shape to
another. For example, figure 7 shows how to change a cup to a doughnut through a series
of small transformations. This measure also satisfies the triangle inequality.
One example of a morphing distance is the Earth Mover’s Distance. The problem
consists of a set of points x and a set of points y. The goal is to find a matrix fij which
tells how much of the mass of each xi goes to each yj , as in figure 8. Mathematically, we
are trying to minimize the function

XX
min cij fij (10)
i∈I j∈J
8 CS273: Handout # 8

where cij is the distance between xi and yj , subject to the contraints

fij ≥ 0 i ∈ I, j ∈ J (11)
X
fij = yj j∈J (12)
i∈I
X
fij ≤ xi i∈I (13)
j∈J

This can be solved via linear programming. The earth mover’s distance is defined as
P P
i∈I j∈J cij fij
EMD(x, y) = P (14)
j∈J yj

5 Topological distance
A final way to measure the distance between two shapes is using topological notions.
One such notion involves the use of writhing numbers, which have previously been used
for DNA. The writhing number counts the number of times that pieces of the protein
chain cross in front or behind one another. It is defined as
1 Z
W = Dw(z)dz (15)
4π S 2
where Dw(z) is the count of intersections in the plane normal to the vector z.

References
[1] C. HuttenLocher, G. Klanderman and W. Rucklidge. Comparing Images Using the
Hausdorff Distance, IEEE Trans. of Pattern Analysis and Machine Intelligence, vol.
15, no. 9, pp. 850-863, 1993.

[2] P.K. Agarwal, H. Edelsbrunner, and Y. Wang. Computing the writhing number of a
polygonal knot. Proceeding Thirteenth Symposium on Discrete Algorithms (SODA),
13:791-799, 2002.