Full text: Proceedings; XXI International Congress for Photogrammetry and Remote Sensing (Part B4-1)

112 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
primary sequence scan, so the crisis of dimension is not avoided 
in the kind of index structures. Then, when the dimension is 
very high, can we give up the idea of partition based index, and 
implement the quick retrieval by reducing the cost of sequence 
scan? 
4.1 VA-File index 
Approximation based indices are implemented through this idea, 
and can receive good performance in the case of high 
dimension. The representative is the VA-File (Roger Weber, 
1998). The main idea is to divide every dimension of data space 
into some segments, form some subspaces or cells, and code 
these cells through bit strings. These cells are represented 
approximately by corresponding bit strings or vector 
approximation. 
vector data 
11 
10 
01 
00 
Figure 4. VA-File ( Roger Weber, 1998) 
From the point of view of approximation, the other category try 
to achieve a more accurate approximate vector and to improve 
the secondary retrieval performance. The kind of indices 
include VA+-File which uses the KL transformation to 
eliminate the correlation of every components, and allocate 
different number of bits to every dimension according to the 
different energy of every dimension after KL transformation 
(Hakan Ferhatosmanoglu, 2000); and IQ index which 
introduces minimum bounding rectangles to VA-File, and 
obtains the data point density in MBRs, and then allocates 
different number of bits to the dimensions of different point 
density (Stefan Berchtold , 2000); and the local polar 
coordinate file (LPC-File) which approximates vector by polar 
coordinates on the partitioned local cells and enhances 
significantly discriminatory power of the approximation 
(Guang-Ho Cha, 2000). From the point of view of the strategy 
of bits distribution and organization of approximate vector 
respectively, these methods improve the primary filtering and 
secondary retrieval efficiency and can achieve better 
performance that VA-File. 
From the analysis, we can find the kind of indices complete the 
quick retrieval by making the cost of sequence scan decrease. 
For the storage space of approximation vector is smaller than 
the primary vector, and then the I/O times needed to access 
decreases sharply. The kind of indices avoid successfully the 
“dimension crisis” phenomenon, and is suitable for the 
application of similarity search where the dimension is very 
large, usually from over dozens dimension to more than one 
hundred dimension. 
The creation of the VA-File is considered as the process of 
building the representation of the approximations. If the number 
of bits to allocate to each dimension is bj, and the total number 
of bits. in vector approximation is 
bj = b , the per- 
dimension partition position can be denoted as ( Pj[0], 
Pj[ 1 ], , Pj[2 bj ]), and the data space is divided into 2 b hyper- 
rectangular cells, each of which can be represented as a unique 
bit-string of length b. Each data point is approximated by the 
bit-string of the cell into which it falls. Figure 4 illustrates this 
for five sample points. 
Similarity search can be performed by scanning the entire 
approximation file, and by excluding the vast majority of 
vectors from the search based only on these approximations 
(Roger Weber, 1998). 
4.2 Improved VA-File index 
In addition, the kind of indices includes still the improved VA- 
File indices. Note that VA-File index all the same uses the 
strategy of two levels “filtering-refine” to complete the high 
dimensional retrieval. Consequently, the improvement of VA- 
File can be done from two aspects. 
In view of this, we classify the improved VA-File into two 
categories according to the strategy. (1) From the perspective of 
the organization method of approximate vector, the first 
category reorganizes approximate vectors according to some 
specific rule and makes the cost to scan the approximate vector 
decrease. The primary filtering performance of the index 
improves for it is easier to find the approximate vector. The 
kind of indices involves sorting approximate vectors based on 
principle component (Cui Jiangtao, 2005) and using R-tree to 
index approximate vector (Cui Jiangtao, 2005) and so on; (2) 
5. DISTANCE BASED INDICES 
Spatial object in spatial data similarity search is usually 
represented as high dimension feature vector, and one of 
important trait is that distance calculation cost or CPU cost is 
expensive. Partition based indices divide the data space 
according to coordinate information without consideration of 
distance cost, and the performance cannot be guaranteed in the 
case of high dimension. And then, is it available to decrease the 
cost of distance calculation by reducing the number of feature 
vectors which are necessary to similarity match calculation, and 
accordingly complete the efficient retrieval of high dimension 
data? 
Distance-based indices utilize the distance properties or 
measure properties in metric space to index high dimension data. 
The formal description of metric space is can be given as 
follows: The set X denotes the universe of valid objects, and U 
is a subset of it, a given set of data objects where we search. 
There is a function d : X X X —> R, which satisfies following 
three axioms. 
(1) Positiveness, 
Vx, y G X , d(x, y) > 0 , when and only when 
X = y , d(x,y) = 0 ; when X * y , d(x,y) > 0 
(2) Symmetry, 
Vx,j> e X ,d(x,y) = d(y,x), 
(3) Triangular inequality, 
\/x,y,z e X, d(x,y) < d(x,z) + d{y,z). 
Then the pair (X, d) is called metric space.
	        
Waiting...

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.