Slashdot Mirror
$6 System-On-A-Chip Mimics Human Vision
Brian McLaughlin writes "This article in TechWeb describes a Visual Perception Processor (costing $6) that can automatically detect objects and track their movement in real time, according to Buereau d'Etudes Vision (BEV). They claim that a full-blown vision processing system/application could be built for less than $50 that rivals current state-of-the-art $10,000 systems. Sounds pretty cool.
" Heck, with my vision, I could tear my eyeballs out and simply use these, at a fraction of the cost of new glasses.
Driving down highway
Blue screen of death suddenly
I crash into tree
The brain-eye system uses layers of parallel-processing neurons that pass the signal through a series of preprocessing steps, resulting in real-time tracking of multiple moving objects within a visual scene.
Get it now? So please guys, don't get any *SMART* ideas and jab your eyes out so you can buy a nice new $6 dollar replacement. (I know no one was actually serious about that, but you guys did still miss the point. Its not the camera thats important, and it would do no good to put xray or a zoom lense on there because you're not looking for the output of what its seeing. Its a TRACKING device. So its not used for monitoring either... geez.) \rant
I am a leaf on the wind. Watch how I soar.
The GVPP's major performance strength over current $10,000 vision systems is its automatic adaptation to varying lighting conditions. Today's vision systems dictate uniform, shadowless illumination, and even next-generation prototype systems, designed to work under "normal" lighting conditions, can be used only from dawn to dusk. The GVPP, on the other hand, adapts to real-time changes in lighting without recalibration, day or night.
Okay, so what they're claiming is that their brand-new, $6 ($50 in total) device can do things which long-standing scientific projects costing $10,000 cannot? Am I the only one who thinks that this sounds somewhat fishy?
The GVPP was invented in 1992, when BEV founder Patric Pirim saw it would be relatively simple for a CMOS chip to implement in hardware the separate contributions of temporal and spatial processing in the brain.
Again don't you think that all of the many computer scientists and neuropsychologists working on machine vision wouldn't have thought of this themselves? I've read a fair bit on the theory of vision processing and pattern recognition and it's a hugely complex subject. And now a small research company has cracked it? I don't think so. If you read the list of things which they say it can be used for it comes across as being a huge gimmick - they seem to have listed everything they could think of that might be worth money.
...if the general-purpose hardware wasn't so stupid. Of the millions of transistors on a modern chip, most of them are wasted in maintaining the illusion of sequential operation, while the OS writers go to considerable trouble to create the illusion of parallel operation.
Furthermore, there are the huge (in terms of transistor count) banks of flip-flops which just sit around most of the time, and the costly layers of cache all working their hardest to maintain the illusion that it is RAM. Meanwhile, software optimizers make sure to access memory sequentially to avoid upsetting this illusion, which would ruin the performance.
You can justify all this nonsense with the argument that software is written for sequential machines with RAM. It's a circular problem. If somebody would just release a cheap massively parallel system, the programmers would learn to use it efficiently.
You can make a complete processor in a few thousand transistors (as this guy has done, though he goes a bit off the deep end...), and you can add a bit (a few K) of high-speed RAM and network them easily enough to make a (dare I say it?) Beowulf cluster on a chip. Each might only run at one tenth the speed of a modern CPU, but you could have hundreds of them for the same cost, giving you bips and gflops for the price of mips.
It would also make the whole design process a lot easier and faster. One simple processor, repeated hundreds or thousands of times. Every advance in production would bring a direct and proportional improvement in performace, with a tiny added design cost. Forget special graphics or sound processors, just plug in more processor banks like you would add memory today and watch your system fly.
C'mon hardware guys, we software guys aren't that stupid! We don't need your illusion of a 386!