International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Since Fant’s algorithm is faster than conventional re-sampling
technique that uses low-pass filtering and sub-sampling, his
algorithm is used as a warping procedure in this study. Also
Fant’s algorithm has one step process comparison with
conventional re-sampling technique (two step). In particularly
conventional method has a complicated process (low pass
filtering or interpolation) in the re-sampling process
3.1 Object Extraction
Since indoor stereo images have a simple back ground, the
object of stereo images can be separated from background.
Processing time is reduced and accurate matching result is
given. This object extraction algorithm consists of six steps :
labeling, top-down merging, bottom-up merging, renumbering,
e o e eov e
small region merging, and object extraction.
3.2 Modified Hierarchical Block Matching Algorithm
In order to reduce noise sensitivity and reach higher efficiency
simultaneously, both the left and right images are low-pass-
filtered and sub-sampled. We propose a modified hierarchical
block matching algorithm shown in Figure 6. Hierarchical block
matching algorithm used in this paper consists of four steps as
follows.
Adaptive Mask Input Images
y
LPF &
Sub- Sampling
>
Bidirection
Check
Interpolation
Block Matching Using
an Adaptive window Warping
7
Next Level Disparity
Figure 6. Modified hierarchical block matching algorithm.
Step ! : Low pass filtering and sub-sampling
Input images are passed a LPF(low pass fitering) and
subsampled.
Step 2 : Block Matching Process
Half-sized input images are used in stereo matching.
Step 3 : Bidirection checking
Disparity map that found in previous process is checked
using the bidirectional constraint.
Step 4 : Interpolation
Remained disparity value is linear-interpolated.
3.2.1 Block Matching Algorithm: First process is a block
matching process. The result of this process is used to generate
reliability function. The matching process is performed in
parallel sequence.
Distance-measure(D(x,y,d)), we used, is Mean-Absolute-
Distance (MAD). Each disparity map is obtained by calculating
the MAD between a pair of stereo images for a particular
window size as Equation (2).
710
N wv
: |o €
D(x, y, d) = arg min | — > lox My y)
de(x,y) Inn NL
Hmm em
-HG d xy y) ;
where d is the window shift in the right image, m and n are the
mask sizes, w is the window size, and 7, and /, are the gray-
levels of the left and right images, respectively. Here we
employ an epipolar constraint to reduce computational cost,
assuming that an epipolar line can be determined first. The
value of d that minimizes the MAD is considered to be the
disparity at each pixel position(x, y).
3.3 Reliability Generation
In order to estimate reliability of matching points, we use
matching constraints: uniqueness and smoothness. Other
constrains are considered in matching process.
3.3.1 Uniqueness Constraint: A given pixel from one
image can match no more than one pixel from the other images.
However, bidirectional check shows that each disparity value
has small difference in Figure 7. And the bidirectional check
distance is defined as
o ID, (x. y) - D, (x Dy(x, y)) ; (3)
The uniqueness condition can be tested for each sampling
position (x,y). The deviation & can be seen as measure of the
perturbation.
right image
left image
Figure 7. The bidirectional check.
Where à describes the difference given by the relation (3). In
order to consider how much input image satisfy the uniqueness
constraints, the following reliability function is employed:
Te md (4)
Where ^»: threshold
Ó : bidirectional check distance
where fp; represents the uniqueness reliability, & is bi-
directional check error, and 75. is minimum threshold.
Ini
3.
ha
thi
Th
ch
sat
ex
Cc
dis
wi
Us
wh
vel
wh
sta
the
dis
hie
