This issue of the Newsletter contains an article on the Engineering Applications Support Environment (EASE) Workstation assessment, carried out last summer. Several machines were assessed, including Apollo, DEC and Sun.
The article on AI support for engineers includes notification of a workshop on AI programming, to be held on 10 March. Also included is a list of courses being run in April and May. These are very popular, so early booking is advised.
Also in the Newsletter are articles on POPLOG and the 4th Alvey Vision Conference, held at Manchester.
There has been a good response to our announcement in the last two issues that a bibliography of AI in engineering and a list of AI tools are available for copying over JANET from uk.ac.ed.aiai. The tool list is continually being updated and has now been split into several separate files. A partial bibliography of AI in control and process engineering has also been added to the pool. The file index contains a list of the files available for copying by ftp.
For example, assuming you use hhcp for ftp, try the following:
Home% hhcp uk.ac.ed.aiai:' <AILIB>index' index transfer authorisation: guest transfer password: <enter your user name here>
Those of you who use the VMS utility transfer, should remember to specify the switch /format=line to ensure successful copying. We have tried to make the file transfer process more robust, file names may be in either upper-case or lower-case, and the files are now purely textual, they don't need to be processed by any text formatters. However, we would remind you to use your user name as the password, that way if there are any problems we can get in touch with you via electronic mail.
Our third workshop, entitled Artificial Intelligence in Process Engineering, took place on 9 December 1988. The workshop consisted of a series of talks, discussion sessions and demonstrations.
There is a report on the workshop, which is available from AIAI. The following papers, a partial bibliography of AI in process engineering and a summary of the discussion that took place are all included in the report:
Our next workshop, which will be on AI Programming, is scheduled for 10 March 1989. This is intended to be of interest and relevance to all engineers. There will be a number of talks covering different paradigms of AI programming. A programming paradigm is a way of approaching programming problems and expressing their solutions. The three main paradigms, which are all commercially available in some form and have been used in engineering applications, that will be covered are:
Following the format of our previous workshop, we would like to limit the number of participants to around thirty. If you are interested in attending, please let us know as soon as possible.
All the courses in the first quarter of 1989 are already fully booked. The course schedule for the second quarter is:
Provisional bookings on these courses may be made now. However, their final availability is subject to the extension of our current contract.
Poplog V13.6 (including Pop-11, Common Lisp, Prolog and ML, with an integrated editor, window manager and masses of on-line documentation, libraries, teaching material etc.) is now available to UK academic users from Sussex University. Commercial users should contact SD-Scicon.
V13.6 provides: improved performance especially in Prolog, Lisp and ML; enhancements to the editor VED, including improved Emacs emulation (with dired); improved on-line documentation and program libraries; support for SunOS 4.0 and VMS 5.0; improved tracing and error handling; new system calls and virtual machine facilities; improved external_load, including segmented heap mechanism so that externally loaded programs can dynamically allocate store at will; a non-copying garbage collector option for limited swap space; support for several new machines.
The Poplog version of the BT/OU KEATS system, ported to Sun Poplog on an Alvey grant, is now available for UK academic sites. It uses the X windows system VlOA at present, but will be converted to X11.
The new machines supporting Poplog are Sequent Symmetry (multiple 80386 with Berkeley UNIX), Sun4 (SPARC), Sun-386i (the "road-runner").
Previously supported machines were VAX(VMS/UNIX), Sun-2, Sun-3, Apollo (with UNIX), Hewlett Packard M68080 Unix workstations, Orion-l/05 (Clipper plus UNIX). Other Poplog ports are likely to follow, especially new 80386-based workstations.
Poplog development is continuing, though at a reduced level owing to the end of Alvey/SERC funding. The version X Windows VII interface is nearing completion. In 1989 a Poplog delivery mechanism will be available enabling small packages to be compiled without including the development environment. Cross-compilation will also be possible.
A catch-up service is available for Poplog users who have not received regular updates.
The Alvey Vision Conference was hosted this year by the University of Manchester and chaired by Dr Chris Taylor. This was the fourth in a series of annual conferences organised by the Alvey Vision Oub to act primarily as a forum for bringing together Alvey supported research groups in the field of Computer Vision and Image Analysis, and for the presentation of their work. The meeting was well attended and drew delegates from throughout the United Kingdom, and a few from overseas.
For the first time in this series of meetings, there was a substantial exhibition of workstations and image processing hardware which was well received by delegates.
Although a number of Alvey projects have now finished, the range of papers was as broad as in previous years, but with some changes in emphasis.
There was a significant number of papers at the meeting on 2-D and 3-D object recognition. On the first morning, speakers from the vision group at the University of Reading described a range of component techniques for matching 3-D models. These included an efficient means for computing the visibility of object features, a method for approximating the orientation of an object from the disposition of 2-D image features, and improvements to an iconic matching procedure described at AVC87.
In other sessions of the conference, a number of new algorithms were presented for detecting 2-D shapes in images, and in particular, a circular arc detector reminiscent of Ballard's strip tree generator for straight line approximation, a novel use of the Hough Transform to detect ellipses, an application of existing speech technology to recognise shapes, and a method for detecting the brain in magnetic resonance images for patient realignment.
The high computational cost of model-matching algorithms was tackled in a single paper describing an implementation on the DAP of Grimson and Lozano-Perez's matching algorithm. Another paper described an information theoretic approach to learning structural relationships for visual purposes, thereby endeavouring to avoid the manual construction of models.
A full session of the meeting was devoted to the work of the Wolfson Image Analysis Unit at the University of Manchester. The goal of this work has been to produce methods for constructing domain specific image analysis systems without the need for a 'vision' expert. thus addressing one of the major problems in the practical application of computer vision, that of re-usability. To this end, the session included an examination of object oriented programming techniques for building vision systems and the use of a commercially available knowledge engineering tool (ART). Other talks were concerned with the detection of grey-level symmetries as cues for model instantiation and a number of boundary detection algorithms; including a non-linear detector that was shown to perform significantly better than a state of the art linear detector.
Another full session of the meeting was concerned with visual guidance systems for autonomous land vehicles. In particular, the Plessey vision group described a system for finding horizontal patches of road surface by:
Three papers from the University of Bristol presented work on the design of hardware and software for a real-time system to guide a vehicle along a road.
Several results and algorithms were presented on surface recovery and representation, including work from Oxford University on the determination of surface orientation from texture, and the recovery of shape information by a moving observer from specularity (including resolution of the convex/concave ambiguity). A method for fitting simple shapes to 3-D line segments was also reported at the meeting - extending work presented at AVC87.
In a comparative study of three stereo matching algorithms, speakers from UCL reported on the advantages of a new algorithm, derived from a standard photogrammetric technique, for matching stereo-pairs from the SPOT IMAGE satellite. They also demonstrated animated terrain reconstructions using the new stereo matcher and standard graphics techniques.
A number of developments in image feature detection and description were also reported at the conference. These included an improved edge detector obtained by fully optimising Canny's performance measure, a method for locating edges more accurately by examining temporal sequences of images, an investigation of the use of grey-scale morphology in feature detection, and a novel method for locating the locus of symmetry points.
One of two papers on optic flow described its application to monitoring the development of eye cataracts. This was one of an number of papers concerned with specific applications of vision technology. Other examples, were papers concerned with the analysis of images of textile fibres for forensic purposes, and the analysis of engineering drawings.
As in previous years there were two invited speakers. The first, David Lowe from the University of British Columbia, described his work on 3-D object recognition and in particular the SCERPO system reported in the recent literature. He concluded that it is possible to build robust 3-D recognition systems without depth reconstruction (i.e.. without the formation of a 2-l/2D sketch) but rather by matching 3-D models directly with 'perceptual groupings' of 2-D image features. The talk finished with the optimistic prediction that within 5 to 10 years we should have low cost systems capable of real-time recognition of any well defined object from any viewpoint.
The influence of Lowe's work was reflected in several papers throughout the conference, including an evaluation of his recognition method and an elegant derivation of constraint equations similar to his own.
The field of computer vision has a history of seeking inspiration from natural vision systems, and the second invited speaker, Dana Ballard from the University of Rochester, continued this tradition by demonstrating the relevance of fixation and eye movements to the design of machine vision systems. The central hypothesis of this talk was that by fixating on a point in space during camera movement and moving the primary frame of reference to this point, and away from the viewer centred frame of reference, many of the problems of early vision become more tractable. For example, image flow about the projection of the fixation point is proportional to relative depth with respect to this 3-D point. Ballard refers to the approach as Animate Vision as opposed to Dead-Eye Vision.
Having made this observation, the rest of the talk was concerned with the non-trivial problem of tracking a fixation point using a sequence of stereo-image pairs. His proposed solution was to track the projection of the fixation point from frame to frame of either the left or right camera using grey-level (without DC component) correlation and then to verge onto the fixation point with the other camera, using the spectrum of the superimposition of left and right views to obtain the stereo disparity. The video showing the performance of this system (in real-time) was impressive.
Selected papers from the conference will appear in a special issue of Image and Vision Computing published by Butterworth Press. The activities of the vision club will continue under the DTI/IED programme, and we look forward to the next conference, which is to be held at the University of Reading from 25 to 28 September 1989.
The prize for the best paper of AVC88 has been awarded to G P Otto, T K W Chau, T Day, and J P Muller of University College London for their joint presentation on stereo matching methods for extracting elevation data from SPOT satellite images. In a clear and interesting report they described a novel approach to an important practical problem. giving quantitative results which included an objective comparison with previously published methods.
The Alvey Vision Club Committee also wish to commend D Charnley and R Blisset of Plessey Research, Roke Manor, who, in a polished presentation, described the application of previously reported structure-from-motion methods to outdoor image sequences acquired by a vehicle-mounted TV camera.
The 45 papers presented were printed in the AVC88 Proceedings (302 pages) available from Margaret Matthews, AVC88 Proceedings Secretary, Dept of Medical Biophysics, University of Manchester.
Vision Club activities will continue under the auspices of the DTI/IED, we look forward to our next Conference to be held in the University of Reading, 25 - 28 September 1989.
The Engineering Applications Support Environment (EASE) Programme differs from the former Common Base Programme (CBP) in that it is oriented towards the Applications software needs of the Community rather than concentrating on the hardware and basic software systems. There are now two categories for hardware: Fully Supported (as before) and Approved. Machines in the Approved category will be expected to run a wide range of EASE software, communicate with the EASE distributed infrastructure and demonstrate good performance. They will not have any specifically funded support (for example, no effort will be available to port software to those machines). However, advice on these machines will be available to grant applicants if required, and it should now be possible to take advantage of special deals without difficulty.
Identifying a machine for Fully Supported status takes a considerable amount of time and effort. Consequently, it is not undertaken very often. The survey which selected the SUN range took place some 3 years ago. It is proposed to keep the Approved list more up to date, using a selection technique which will provide the required information with sufficient accuracy without taking a lot of time. The Approved list will form the basis of the shortlist from which candidates for Full Support will be selected.
It was decided to assess Workstations as the first such exercise, mainly because of the changes in marketplace since the last assessment, and because more is known about these machines. This is a summary of the report which presents the results. A more detailed version is available on request, although it is not possible to make the original report generally available (since it contains confidential data).
The assessment was carried out by producing a questionnaire, intended to elicit the desired technical information from the suppliers, and by running benchmarks on machines loaned for the occasion. These tests were designed to give a rough guide to the performance of the machines, and the results take this into account. Much more stringent tests would be needed if a full evaluation were to be undertaken. The final report was presented to the Computing Facilities Technical Advisory Group (CFTAG) in September.
The following table (1) gives details of the Workstations which were assessed. Machines based on the Intel 80386 chip and the more powerful graphics workstations were excluded from this assessment (they will be handled separately). At first, DEC suggested that the DEC 2000 be tested However, this machine proved to have rather less power than the others, and tests were then run on the DEC 3000 range. Both sets of results are included in the paper.
All the suppliers were most co-operative in answering the questions and providing machines for benchmarking, which greatly assisted in the production of this report.
The price details refer to the list price of the machine actually benchmarked, without reference to any form of discount Screen size indicates the maximum horizontal addressability on a 19in monitor.
| Make | Machine Range |
I/O Bus | Heat (KW) |
Noise (dB) |
Screen Size |
Machine Tested |
Memory (MB) |
Disc (MB) |
Price Bracket |
|---|---|---|---|---|---|---|---|---|---|
| Apollo | DN3000 | PC | 0.5 | 55 | 1280 | DN3500 | 16 | 348 | Medium |
| DEC | 2000 | None | 0.3 | 40 | 1024 | 2000 | 6 | 154 | Low |
| DEC | 3000 | Q | 1024 | 3500 | 16 | 154 | High | ||
| HP | 9000 | VME | 0.3 | 50 | 1280 | 9000 | 8 | 160 | Medium |
| HP | /300 | /350 | |||||||
| IBM | 6150 | PC | 0.3 | 51 | 1024 | /125 | 8 | 180 | Medium |
| SUN | 3 | None | 0.8 | 45 | 1152 | 3/60 | 4 | 141 | Low |
| SUN | 4 | VME | 0.8 | 1152 | 4/110 | 8 | 327 | Medium | |
| Tektronix | 4300 | VME | 1.3 | 50 | 1280 | 4335 | 12 | 110 | High |
There is not space to report much detail of the results of the exercise. The following information is a small subset only, to give the reader a feel for the style of report. In particular, note the WHICH Report style tables which provide information in a form which does not require detailed evaluation of the benchmark results. Cost effectiveness is not considered explicitly, since the actual price paid for a particular product tends to be very different from the quoted list price in any particular instance.
All results are given relative to those of the SUN 3/60.
Key to tables:
Code Range of ratios to SUN 3/60 Blank 0.9-1.2 + 1.2-1.5 ++ 1.5-2 +++ 2-3 ++++ >3 - 0.7-0.9 -- 0.5-0.7 --- 0.3-0.5 ---- <0.3 * No meaningful result
Table 2 gives relative values for basic arithmetic operations, function calls, system calls and disc I/O (all called from C).
| Integer | FP | Function Calls |
System Calls |
I/O | |
|---|---|---|---|---|---|
| Apollo | ++ | * | ++ | +++ | |
| DEC2000 | --- | -- | ---- | ||
| DEC3500 | + | ++ | + | ||
| HP | - | + | + | ||
| IBM | - | + | -- | + | - |
| SUN4 | - | + | ++ | ++ | +++ |
| Tektronix | + | ++ | ++ |
The following tables (3a and 3b) give relative values for basic arithmetic operations, function calls, and input/output from Fortran and Pascal programs (it was not possible to run the Pascal tests on the Tektronix).
Compilers used with the DEC 2000 were the standard Unix ones.
| Integer | FP | Function Calls |
I/O | |
|---|---|---|---|---|
| Apollo | +++ | * | ||
| DEC2000 | ---- | + | ---- | ---- |
| DEC3500 | ++++ | --- | ||
| HP | +++ | + | ||
| IBM | - | ++ | -- | - |
| SUN4 | --- | +++ | +++ | |
| Tektronix | + | + | -- |
| Integer | FP | Function Calls |
I/O | |
|---|---|---|---|---|
| Apollo | +++ | * | ||
| DEC2000 | -- | - | ---- | -- |
| DEC3500 | ++ | ++++ | ||
| HP | +++ | * | ||
| IBM | -- | -- | --- | - |
| SUN4 | ++++ | ++++ | +++ | |
| Tektronix | * | * | * | * |
Table 4 gives relative values for the performance of real applications in the following selected areas:
The finite element program would not run on the Tektronix.
| IBM 6150 |
Tk 4335 |
SUN4 110 |
DEC 2000 |
DEC 3500 |
HP9000 350 |
Apollo 3500 |
|
|---|---|---|---|---|---|---|---|
| Computation | - | ++ | -- | ++ | + | ||
| Finite Element | * | + | -- | ++ | ++ | ||
| Image Processing | - | - | -- | ++ | + | ||
| IKBS | -- | - | -- | --- | -- | + | - |
The benchmarks tested NFS functionality and performance when using an EASE Central Server. In general, all the machines performed adequately. However, IBM do not at present have a NFS product and there are problems over functionality and performance with the current released version of NFS from Apollo. Both suppliers are in the process of introducing new versions of their systems which will overcome these difficulties.
All provide GKS with Fortran bindings. Apollo and DEC are at level 2B; the rest are at level 2C. Only IBM can provide GKS-3D (as a third-party product). SUN, IBM, Tektronix and Apollo provide PHIGS, with Fortran bindings.
It was only possible to test X.11 on HP and DEC 3500 (all the rest could only offer X.10). The version run for comparison on the SUN 3/60 was an unoptimised direct port of the MIT code, so its performance was low. The results of the tests showed that the HP system performance was comparable to that of the SUN 3/60, whereas the DEC 3500 performance showed a significant improvement.
Table 5 presented in this section is an attempt to summarise the general characteristics and performance of each machine. This has been done by assessing each characteristic for each machine on a scale of * (poor) to ***** (excellent). The assessments are non-competitive, and do not represent a ranking of workstations against each other.
| SUN 3/60 |
IBM 6150 |
Tk 4335 |
SUN4 110 |
Apollo DN3500 |
DEC 2000 |
DEC 3500 |
HP9000 350 |
|
|---|---|---|---|---|---|---|---|---|
| Basic System | ||||||||
| Hardware | *** | *** | *** | **** | **** | ** | ** | *** |
| Software | *** | *** | *** | *** | *** | *** | *** | *** |
| Communications | **** | * | *** | **** | ** | *** | *** | **** |
| Graphics | **** | **** | **** | **** | ** | *** | *** | ** |
| X-Windows | ** | * | * | * | * | * | **** | *** |
| Languages | ||||||||
| Standard | *** | ** | *** | *** | **** | **** | **** | *** |
| Additional | ***** | ** | * | **** | *** | *** | *** | ** |
| Applications | ||||||||
| Availability | **** | *** | ** | **** | **** | *** | *** | ** |
| Performance | ||||||||
| Basic System | *** | *** | *** | **** | **** | ** | **** | **** |
| Applications | ** | *** | *** | **** | **** | ** | **** | **** |
The IBM Communications rating reflects the lack of NFS at present. The Apollo Communications rating reflects the state of the system at the time the tests were carried out. The ratings for all systems except DEC 3500 and HP on XWindows reflects the non-availability of X.11.
There have been a number of significant changes to some of the systems since the tests were undertaken. New versions of Operating Systems have appeared, providing performance enhancements. Unfortunately, it is not practical to be continually updating a report such as this, and so readers must be aware that certain sections may well no longer be accurate.
In general, using a questionnaire and benchmarking a single system from each series has succeeded in obtaining a significant amount of information from suppliers whilst using limited resources. It does not appear to have been too great a burden for the Supplier (so that the method may be used again). However, experience has shown that the questionnaire needs revision (to remove redundant questions, clarify those open to misinterpretation, etc.). In addition, more Applications oriented benchmarks are required.
The original report was presented to CFTAG in September. As well as these results, the Group also considered the availability of manpower required to maintain the Approved list, and the relative visibility of the various machine ranges in the academic community. On the strength of all the evidence, the recommendations were made to the Computing Facilities Committee (CFC) who have agreed the following:
CFTAG also recommended that the IBM and Apollo range should be added to the Approved list as soon as the remaining questions relating to NFS have been answered and the products have been released.
It is intended to update this exercise in the summer of 1989, having undertaken an assessment of 80386 machines beforehand.
Network Services Seminar, 23 February 1989 Lecture Theatre, Bldg R22 Rutherford Appleton Laboratory
Workshop on Data Exchange, 11-12 April 1989 University of Leeds
