243 
mple the 
of 1024 x 
:arry out a 
r example 
ire targets 
The performance of workstations can be improved with the use of additional boards for fast 
processing. Gruen (1989) has also reviewed developments in this area but some additional 
comments can be made. A component of particular interest to photogrammetrists is the TAAC-1 
accelerator from Sun. This has an 8Mb image memory and can transfer data sequentially at 50Mb 
sec "1; it can be added onto a Sun workstation giving greatly increase performance for handling 
digital imagery. The TAAC-1 needs a Sun with a VME bus which makes the total system (Sun 
plus TAAC-1) expensive. A TAAC-2 is promised which will fit onto Suns S-bus and reduce the 
cost of the system by about 50%. 
The Inmos transputer has already shown considerable potential for fast processing when used 
on-line or off-line with DPWs. The Kern DSP uses transputers for tasks needed intensive 
computing fast, such as image correlation. Transputers are also used for image correlation in 
systems developed at EPFL (Kolbl, Workshop on fast processing) and UCL (Muller et al, 1988). 
Other special components have been considered for photogrammetric use. At the Workshop on 
hardware and software for fast image processing Spoer described a system based on RAM which 
has high data rate and flexibility of data access allowing quick motion display and fast pan and 
scroll in large arrays. The system, known as ISP 1024, is ideal for image sequence processing, 
interactive editing of raster images, computer animation and medical image processing. Its 
disdvantage is cost being up to five times more expensive than disk based systems. Gunzinger 
described a synchonous data flow machine (SYDAMA) for real time image processing. It is a 
programmable hard wired system which is best suited for low level image processing. Bench 
mark testing on a correlation algorithm have indicared speed ups of 160 times over FORTRAN on a 
VAX 11/780, 100 times over assembler code on a 68020 CPU and 50 times over a transputer. 
Technical specification are given in table 5. Other alternative, but well known architectures which 
could be used in a DPW are the fine grained microprocesor SIMD machines. 
of a DPW 
shows the 
Table 5. Technical Specification for a synchonous data flow machine (after Gunzinger) 
RCE 
Proceesing rate 50 frames /sec at 400 x 300 (256 x 256) pixels resolution 
and 7.5 MHz sampling rate 
ion 
:h 
Processing elements 256 
Communication network 12 circulating 8 bit busses, 7.5MHz data rate 
er 
Matching unit data synchonisation in between 8 picture lines 
Processing elements LUT; A/D and D/A converter; binary; statistics; convolver etc. 
Host computer IBM-AT or compatible 
System software Compiler, configurator; loader; monitor; debugger. 
0 image, 
and thus 
require a 
Storage on 
may need 
want data 
ige needs 
eroding 2 
1 from the 
d bring in 
It is apparent that the trend in architecture for fast image processing is towards modular machines 
which use a standard workstation as a host which will use fast add pon such as the TAAC and 
interact with hybrid systems such as programmable hard wired devices and transputers. Figure 1 
is a scheme for such a system which was proposed at the London workshop. 
Requirements for stereo viewing 
Several methods of stereoviwing are possible. The split screen method is most closely related to 
conventional photogrammetric viewing. It must be possible to move the two parts of the screen 
independently and the optical system must allow for independent image rotation as it is not 
desirable to resample the image to achieve this digitally. A digital zoom is sufficient if both images 
are at the same scale initially. The disadvantage of the split screen method is the viewing optics, 
which must be able to be removed, and the reduced size of the image. A dual screen system is also 
possible but requires an even more cumbersome viewing and display configuration.