Radu Banabic
11-21-2008, 01:47 AM
Hi.
I'm quite eager to see if the DSMC will really be able to do what Red says it will. If so, it will really bring photography (still/movie) to a new level.
I would really like to see a module allowing the user to write/use their own code in the camera. I can think of quite a few uses for such a module, first in mind being computer vision. It wouldn't hurt to be able to control not only the output signal, but also some internal mechanisms like focusing (customized tracking) or zooming. I am certain that the automobile industry will be interested in the processing power of such a camera (see Autoliv pedestrian detection). Not only this, but it might even improve the processing workflow, because you could do some of the processing online and decrease post-processing (heck, maybe the software companies might even start to write code for dedicated in-camera hardware, rather for the slow, multi-purpose PCs).
It doesn't necessarily mean that the user should write actual software, because this slow down the pipeline (maybe several parallel processors would help, if you can "suck" parallel feeds from different parts of the sensor). It could be as trivial as proividing some hardwired (and fast) predefined operations / convolution kernels such as Gaussian, edge detect, corner detect etc and simply allowing the user to choose which and when to activate.
Anyway, good luck with the camera. I still find 28K hard to believe.
Radu
I'm quite eager to see if the DSMC will really be able to do what Red says it will. If so, it will really bring photography (still/movie) to a new level.
I would really like to see a module allowing the user to write/use their own code in the camera. I can think of quite a few uses for such a module, first in mind being computer vision. It wouldn't hurt to be able to control not only the output signal, but also some internal mechanisms like focusing (customized tracking) or zooming. I am certain that the automobile industry will be interested in the processing power of such a camera (see Autoliv pedestrian detection). Not only this, but it might even improve the processing workflow, because you could do some of the processing online and decrease post-processing (heck, maybe the software companies might even start to write code for dedicated in-camera hardware, rather for the slow, multi-purpose PCs).
It doesn't necessarily mean that the user should write actual software, because this slow down the pipeline (maybe several parallel processors would help, if you can "suck" parallel feeds from different parts of the sensor). It could be as trivial as proividing some hardwired (and fast) predefined operations / convolution kernels such as Gaussian, edge detect, corner detect etc and simply allowing the user to choose which and when to activate.
Anyway, good luck with the camera. I still find 28K hard to believe.
Radu