ole de- 
nosaic, 
Relational matching for Stereopsis 
Ofer Zilberstein 
2 Survey of Israel 
the ra 1, Lincoln St. 
ctions, Tel-Aviv, Israel 
ıture a 
Radar 
> areas KEY WORDS: relational matching, tree matching 
xtends Abstract 
sed for ; ; : ; : 
This paper presents the design and implementation of relational matching for solving the 
CITOT correspondence problem. Application such as large scale urban areas or close range scenes 
house- with large depth range relative to the base/ height ratio often pose unsurmountable problems 
an be to existing matching methods. Relational matching can easier cope with geometrical dis- 
e tortions and it is less sensitive to the presence of occlusions, foreshortening and breaklines. 
un the Relational matching is performed by, first, representing epipolar scan lines as trees, and 
: alistic then searching for the minimum distance between the two structures. To demonstrate the 
‘radar flexibility of the matching scheme two different input signals were used. The first input is 
€ an gray levels epipolar scan line, used for matching gray level images, and the second input 
S will involves convolution values, generated by the LoG operator, used for matching, indirectly, 
zero-crossings. The results show that the matching scheme copes successfully with large 
geometric distortions without introducing special constraints or tuning the algorithm. 
metric Ss : : 
ect The most prominent image matching problems are relief 
ipn distortions, occlusions, discontinuities in the surface (break- 
ersity, 1 INTRODUCTION lines) and non-linear radiometric differences. Applications 
Most photogrammetric processes involve two or more pho- Sa as oe Scale Sn amas 67 Cote Tonge scenes Sith a 
on of tographs. One of the most fundamental tasks in photogram- arge depth range relative to t e base/ height fatio often pose 
1077- metry is to identify and to measure a feature of the object unsur pn problems to ns matching methods. 
space in all overlapping photographs. In photogrammetry, Re ational matching I5 new in o Otogrammctry eIn com: 
the process of finding conjugate features in two or more im- T RANT it is the ACER bey Lite Yision for 
mann, ages is commonly referred to as the image matching problem. mate um T eatüres vr a mode ase (object Iasi) 
Borne The image matching problem can be described as comparing [4]. Re ational maie mg con Cane with ER er 
ch. f a specific feature with a set of other features and selecting Hons etween «he festures to he Compared Much easier, X 
: 2 ; m : : is less sensitive to the problems mentioned earlier. Hence it 
p 
-161. the ‘best’ candidate based on criteria such as shape, intensi- hould b bust method f line th 
actical ty values, etc. In traditional photogrammetry this problem Shou'd pe 9 more FoDusu method. or so Ving LUS Correspon- 
; : : : ; dence problem. This paper focuses on relational matching, 
: is solved by an operator who identifies conjugate points by : EE 
lation. fusing the overlapping photographs to form a stereo mod- explores its potential in digital photogrammetry and demon- 
Vol. el. The human visual ability to solve the cortespondence strates the usefulness in the surface reconstruction problem. 
problem is unsurpassed and performed in real-time, without 
nn, P. conscious effort. Not only does the human visual system for- 2 MATCHING TECHNIQUES 
j of m a stereo model but it interprets the 3-D model and stores 
h f a highly symbolic description which is more useful than the Detailed reviews of image matching techniques in computer 
A original light intensities to draw conclusions and to properly vision can be found in Barnard and Fischler [5] covering 
act to what one is looking at. the period from mid-70's until 1981. Dhond and Aggarwal 
) pro- How the human visual system accomplishes this feat is [10] review the topic of ‘structure from stereo’ from 1981 to 
osium largely unknown. This is particularly true for the higher 1989. Li [25), Hannah [16], and Doorn et.al. [11] review 
8 p 8 
ity of 
FR.S. 
\erial 
ESA 
1985: 
fetric 
-209 
visual processes such as image understanding and object 
recognition. The correspondence problem is considered an 
early visual process. That is, fusing two images to a 3-D 
model (stereopsis) is presumably being performed without 
a priori knowledge about the scene. 
711 
image matching in digital photogrammetry. 
In general, three criteria characterize matching techniques 
(1) The selection of features and relationships to be matc- 
ed. Features can be in the form of patches extracted