ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002 
  
automatic matching of highly deviated landmark-less aerial 
images (Nishikawa, 2001), which provides a fundamental 
function in realizing automated land cover change monitoring. 
In the present paper, basic concepts are discussed in section 2 
and applications are presented in section 3. For the sake of 
simplicity, primary discussions are limited to two-dimensional 
and two-level hierarchical objects, although extensions to more 
general objects are possible. 
  
Figure 1. Ill-configured objects (ICOs) 
(a) School; (b) segmented image; (c) and (d) layout change; (e) layout 
and component change; and (f) size change. An example representation 
of bis, 
O.attrib = school, O.area = (x, y, w, h), O.area.size = (w, h), 
O. area. base = (x, y), O.subs = {B, G, W}. We should prepare different 
O.layout for each of b, c, d, and e. For f, O.area.size should be 
changed also. 
2. BASIC CONCEPTS 
2.1 Ill-configured Object 
Assume O denotes a geometrical object contained in an image 
I which can be approximately represented by a tuple: 
O = (O.attrib, O.area, O. subs, O.layout), 
where O.attrib is the attribute, such as the name, of O. 
O.area = (x, y,w,h) denotes the minimum bounding rectangle 
(MBR) of O (Samet, 1993). Occasionally, we use O.area.size 
and O.area.base to represent the size (w, ^) and the base point 
(x, y) (north-west corner) of the MBR in 7, respectively. 
A - 388 
O.subs denotes the set of components (sub-objects) of O. 
O.layout is the layout of O.subs. 
In the case in which O is a terminal object (having no sub- 
objects), O.subs — O.layout - 9, where $ denotes a null set. As 
we are concentrating on two-level hierarchical objects, O.subs 
are composed of terminal objects. We refer to O as well- 
configured (or not ill-configured) if O can be represented by a 
small set of such models. For example, trademarks and human 
faces are well-configured objects. However, as shown in 
Figure 1, several objects of concern to us are not well- 
configured, and these objects are referred to as ill-configured 
object (ICO). As discussed before, traditional representations 
such as graphs and others are not useful for ICOs because the 
number of models required in representing an ICO becomes 
very large (as shown later) and/or the cost of model 
application (matching) to real data becomes very large. 
(d) 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 2. Neighbour set representation (NSR) 
(a) School image; (b), (c), and (d) neighbor set of the building, water, 
and ground, respectively, when their placements are fixed (non-ICO); 
(e), (f), (g) neighbor set when the building is fixed at left but other 
components can move (ICO); (h), (1) and (j) neighbor set when all three 
components can move freely (ICO). In this example, model size 
(=| O.subs |) = 3. 
2.2 Neighbor Set Representation 
First, we introduce the neighbor set representation of an object 
O, denoted as NSR(O). 
Definition 1: NSR(O), the neighbor set representation of an 
object of O, is defined as follows: 
NSR(O) = (O.attrib, O.area, O.subs, O.nset), 
Note that, the term O./ayout in the original definition of O is 
changed to O.nset. O.nset is a neighbor set of O. In other