« o e 3rd Step 
oe 2nd Step 
$ n3 St, Step 
000800 00800000008 
9 ISO iis Oe 9o 
225 ee © 6 © 0900909 999 © © 
0:02,02 8490292098 
ees dg giao 
MEN papse iie 
Saint o elo Nie roo 
Pu 
y 
| 38— 
  
225 =! 
X 
Fig. 1 Disposition of grid points for 3 steps 
3 ; ë 
Conjugate Point 
Lace 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
zZ ie 
by x 
1 1 y. À 
e A 
- v 
1 : I 
Draht 
Y oY 
i Lo) = 
vipat 
ry 
4); 1 . A4 
4$. 34 1 
a ani audi 
ATE AT eT 
y Frog Sn gat 
T 42 359 . 
A. 
Fig. 2 Rearrangement of pixel arrays of a right patch’ 
The details are discussed in section 4. 
(8)The processes (2)-(7) are repeated for every step. After that, grid 
points located in the margin of the left patch are looked on unreliable 
and discarded. 
3. Stable matching with narrow-band-pass filters 
We discuss here band-pass-filtering that leads to stable matching in 
area correlation. Indeed this problem is in itself very significant, but 
we specifically need stable matching performance so that the method of 
finding occlusions, detailed in section 4, may really come into effect. 
For comparison let us begin with the following primitive filter, which 
can be derived from very natural intuition, and in fact has been used in 
our former system, but does not give good matching stability. 
Since the highest (angular) frequency is 7. in digital images, the 
patches I, and I, matched in the 1st and 2nd steps are low-pass-filtered 
beforehand such that the higher frequencies than œ, =1/4Æ and @, = 1/2 Z 
are cut off respectively. Further they are resampled every 4 and 2 
pixels to save memories according to the sampling theorem. We refer to 
the finally produced patches as the 'reduced' ones hereafter. 
- 319 -