Image Processing Reference
In-Depth Information
accurate version of this model by assuming a different affine transform, and thus a different
d
, for
each facet of the mesh. This approximation neglects depth variation across individual facets, rather
than across the whole surface. Under this assumption, the projection of a 3D point
q
i
lying on facet
f
of the mesh can be expressed as
d
f
u
i
v
i
A
I
2
×
2
0
q
i
+
0
,
=
(3.2)
where
d
f
accounts for the average depth of facet
f
. Here, without loss of generality, we expressed
the 3D point in the camera referential, and therefore replaced
R
with the first two rows of the
3
3 identity matrix and the translation with a zero vector. Note that this does not prevent us from
accounting for camera motion. It simply means that it will be interpreted as a rigid motion of the
object of interest.
Under the full perspective model, the projection of a 3D point
q
i
×
is written as
⎡
⎤
u
i
v
i
1
A
I
3
×
3
q
i
+
0
,
⎣
⎦
=
d
i
(3.3)
where the matrix of internal camera parameters
A
is now a 3
×
3 matrix, and each point
i
has a
different depth factor
d
i
.
3.2
3D-TO-2DCORRESPONDENCES
Detecting feature points in images has received enormous attention in the Computer Vision commu-
nity. For most template-based approaches, feature points are typically detected with either the SIFT
keypoints detector
Lowe
[
2004
] or Harris's corners detector
Harris and Stephens
[
1988
]. Once fea-
ture points have been detected in two images, they need to be matched to produce correspondences.
When using SIFT, this can be done by a simple dot-product between specific vector representa-
tions of the feature points. For Harris's corners, methods based on Randomized Trees have proved
efficient
Lepetit and Fua
[
2006
]. From a large set of views obtained by applying random affine trans-
formations to a reference image, a tree that models the relationships between neighboring keypoints
is built. Each leaf-node of the tree then corresponds to a specific keypoint, and matching can be
done by dropping the feature points of a new image down the tree.
Another way to establish correspondences is to first tackle the 2D non-rigid image regis-
tration problem. Non-rigid image registration aims at finding a transformation between two im-
ages of the same surface undergoing different deformations. Different parameterizations have been
proposed to represent the transformation, such as RBFs
Bartoli and Zisserman
[
2004
], thin-plate
splines
Bookstein
[
1989
], or 2D meshes
Pilet
et al.
[
2008
]. Since the resulting warp is defined over
the entire image, discrete correespondences can then be obtained by sampling it. Note that, while
2D non-rigid registration can be thought of as related to 3D non-rigid reconstruction, we believe
Search WWH ::
Custom Search