The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
of planes having the baseline as axis (the baseline is the line 
joining the camera centres). This geometry is usually motivated 
by considering the search for corresponding points in stereo 
matching. 
There are three degrees of rotational freedom for each camera 
in total. However, a rotation around the optical axis of a camera 
will not change the visual data, just the orientation of the image 
of that camera. Furthermore, in order to keep the cameras in 
fixation, there will not be any relative tilt between the cameras 
and a joint tilt around the baseline, will not change the nature of 
the problem. 
2.2 Image pyramid 
An image pyramid is a series of pixel arrays that represent the 
same image at successively reduced resolutions. Image 
pyramids are used extensively in multi-resolution image 
processing and analysis. In addition, they have also been used 
in data compression of images and design of computer 
architecture. The generation of an image pyramid for a given 
image can be viewed as the application of a series of low-pass 
filters with successively narrower bandwidths. The input image 
is fed through this series of filters, and the output of each stage 
corresponds to the reduced-resolution image at each level of the 
pyramid. The design of a good pyramid generation process 
becomes the problem of finding an efficient and effective way 
to perform such filtering operations. Convolution with limited- 
size kernels represents fast implementations of low-pass filters. 
The common approaches for generating pyramid data structure 
include Gaussian pyramid, Facet pyramid and Wavelet pyramid. 
In this paper, the Laplacian pyramid is adopted. 
2.3 Memory pool strategy 
There will be some additional costs in the course of using the 
default memory management function new/delete or 
malloc/free in the allocation and release memory stack. 
In the receiving system to a certain size memory allocation 
request, first of all, find the maintenance of internal memory 
free block list, and some algorithms are needed to find suitable 
free memory block size. If the free memory block is larger than 
standard, it is necessary to cut distribution of spare parts and 
smaller pieces. The system then updates free memory block, 
completes a memory allocation. Similarly, in the release of 
memory, the system adds the released memory to the free 
memory block table. If it is possible, they can block the 
adjacent idle merged into larger free block. 
The default memory management function also takes into 
account the multi-threaded applications, which increase 
spending because of the lock for each allocation and release 
memory. If the applications allocate and release memory 
frequently in the stack, the performance will be descended. And 
a large number of fragments of memory are generated, which 
reduced memory utilization. The default memory allocation and 
release of natural algorithm also takes into account the 
performance, but these memory management algorithms are 
used for meeting more complex, more extensive work. For a 
specific application, a customized memory pool management 
algorithm would be offered better performance. 
Because the images obtained from the pyramid image database 
are fixed in size, so a fixed memory pool scheduling algorithms 
is adopted. Fixed memory pool consists of a range of fixed-size 
memory block components, each memory block contain a fixed 
number and size of the memory cell. 
Figure 1. Fixed memory pool 
As the figure 1 shows, the total memory pool includes four 
memory blocks. In the first generation of memory pool, only a 
memory block is applied, and the pointer is returned as the head 
pointer to the whole memory pool. With the applications 
demand more memory, the memory pool is dynamic expand 
and once again to apply for a new memory blocks, and all 
memory block are linked together through the pointer. 
For an operating system, it has been allocated the same size of 
four memory blocks to the application. The distribution of 
speed relatively quickly because of the fixed size. For the 
application procedures, the pool opened its memory a certain 
size, but also the internal memory pool of surplus space. 
Take the fourth memory block as example, which contains part 
of block head information of memory pool and three memory 
modules with same size. Unit 1 and Unit 3 is idle, Unit 2 has 
been allocated. When the application needs one memory 
module which allocated from the memory pool, only a simple 
traversal of all the memory pool block head information, and 
rapid positioned to the memory block which have free unit. 
Then according to the block head information, directly 
positioned to the first vacant unit address and return the address 
and mark the next free memory modules. When the application 
releases one memory pool unit, mark this unit vacant in the 
corresponding block head information. 
Compared with the memory management system, the operation 
of memory pool very quickly, performance optimization in the 
following main advantages. 
(1) For special circumstances, such as the need for frequent 
distribution and release of fixed-size memory object, does not 
require complex distribution algorithms and multi-threaded 
protection. No additional overhead in maintaining the free 
memory form. 
(2) Since the opening up a certain amount of memory space for 
memory pool as a block, thereby to some extent improved 
procedures localised, and enhance the performance. 
(3) Relatively easy to control the page border alignment and 
memory byte alignment, which avoid the memory fragment. 
3. DYNAMIC EPIPOLAR REARRANGEMENT 
This paper possess an advantage of generating epipolar images 
from the original images pyramid dynamically, so in this