Graphics Reference
In-Depth Information
n(p)
n(p)
c(p)
c(p')
c(p)
patch p
images V(p)
image
(a)
(b)
Figure 8.26.
(a) To compute a 3D patch's score in PMVS, we sample it on a regular grid in 3D,
project these samples to points in each image in which the patch might be visible, and compute
the normalized cross-correlation of blocks of intensities around the projected locations. (b) If a
cell of the coarse image grid (dark gray) has no corresponding patch, we hypothesize a the center
of a new patch
p
by intersecting the viewing ray through the cell with the plane corresponding
to a nearby patch
p
from an adjacent cell (light gray).
(a)
(b)
Figure 8.27.
(a) Six of sixteen input images for a multi-view stereo algorithm. (b) Two views of
the 3D result of PMVS (after creating a triangle mesh from the estimated points).
The next step is to expand the patches generated for high-quality feature
matches into regions where no good features were found. This is accomplished by
finding a cell of the coarse grid in some image that has no corresponding patch but
has a neighbor cell with a well-estimated patch
p
. We simply create a new patch
p
for
the patchless cell, estimating
c
p
)
as the intersection of the viewing raywith the plane
containing the neighbor's 3D patch, and initializing
n
(
p
)
=
p
)
=
.
The process of refining the patch parameters then continues as shown previously. If
thefit is poor (e.g.,
p
isnot visible inenough images, or straddles adepthdiscontinuity
in the scene) the new patch is rejected.
(
n
(
p
)
and
V
(
V
(
p
)