Hi Erik,
thank you very much for sharing your experiments.
I totaly agree with you the DVS128 camera has a low image quality. But the new prototype with active pixels (Dynamic and active pixels sensors - DAVIS) and higher resolution (240 x 180 APS) allows us to get better results as your DVS128.
I am very interested in your comparision results with Nupic, which will hopefully give you comparable results but with standard PC, without use of any supercomputer.
What do you think?
Best regards
Binh
From: Erik Rehn <[email protected]>
To: [email protected]
Subject: Re: [nupic-discuss] NuPic with Neuromorphic camera?
Message-ID: <[email protected]>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Hi,
I've done some work with a DVS128 sensor. I wrote an interface between
it and a large scale simulator running on a Cray XE6, and tested it with
an autoassociative recurrent attractor neural network. A report of
what I did can be found here:
http://www.diva-portal.org/smash/get/diva2:547770/FULLTEXT01.pdf
Although I think the concept of a silicon retinas is interesting I was
quite disappointed on the DVS. It is was hard to make it do anything
useful. How "neuromorphic" it is can be discussed. It solely relies on
temporal contrast, meaning that unlike the vertebrate visual system,
which employs spatial edge detection and spatial contrast adaptation, it
only detects local illuminance changes at single pixels. As you say,
this has the effect that a static scene will not produce any output
events. The sensor has either to be moved or pointed at a moving scene
to generate any output other than noise.
For me the DVS seems most useful for applications were a high-speed
camera would normally be needed. It has very low latency. However, in
situations where one can tolerate some latency it's better to use a
conventional camera and convert the image to a spike/address-event
representation in software. Then you have more control, and much higher
resolution.
Best,
Erik
To: [email protected]
Subject: Re: [nupic-discuss] NuPic with Neuromorphic camera?
Message-ID: <[email protected]>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Hi,
I've done some work with a DVS128 sensor. I wrote an interface between
it and a large scale simulator running on a Cray XE6, and tested it with
an autoassociative recurrent attractor neural network. A report of
what I did can be found here:
http://www.diva-portal.org/smash/get/diva2:547770/FULLTEXT01.pdf
Although I think the concept of a silicon retinas is interesting I was
quite disappointed on the DVS. It is was hard to make it do anything
useful. How "neuromorphic" it is can be discussed. It solely relies on
temporal contrast, meaning that unlike the vertebrate visual system,
which employs spatial edge detection and spatial contrast adaptation, it
only detects local illuminance changes at single pixels. As you say,
this has the effect that a static scene will not produce any output
events. The sensor has either to be moved or pointed at a moving scene
to generate any output other than noise.
For me the DVS seems most useful for applications were a high-speed
camera would normally be needed. It has very low latency. However, in
situations where one can tolerate some latency it's better to use a
conventional camera and convert the image to a spike/address-event
representation in software. Then you have more control, and much higher
resolution.
Best,
Erik
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
