Thursday, April 11, 2019

Hillhouse Renamed to CelePixel and Relocated to Shanghai

Hillhouse Technology Singapore has been renamed to CelePixel Technology and relocated to Shanghai, China. The company develops neuromorphic event-driven sensor and has filed for 7 US patents:

"In 1989, Carver Mead, US computer scientist, a founder of Moore’s law and VLSI, created the concept of Neuromorphic Engineering.

In 1990s, his students Misha Mahowald and Kwabena Boahen developed the first Retinomorphic sensor based on Address Event Representation. Subsequently, a number of scientific institutions started to research on Retinomorphic sensors.

Standing on the shoulders of giants, CelePixel has gone further in technological innovations and explorations, to take the cutting-edge underlying technology to forefront of commercial applications.
"


The company has won Audi Innovation Lab Award:

12 comments:

  1. 5ns sampling period? This is not possible. And it would correspond to an equivalent of 200M frames nut 200K.

    ReplyDelete
    Replies
    1. I can't read most of the links as I don't speak Chinese. However, 5 ns are very well more than possible - even for focal plane arrays. Just look at Prof. Goji Etoh who works on >Gfps sensors - in Silicon. Then you have Streak cameras and alike that can go even beyond that (check out e.g. the MIT work). And then please keep in mind that this is not a regular focal plane array but an event driven imager. Only the pixel receiving significant change in radiance have to be read, which makes this much easier. A lot of things are possible these days. Question is what price/trade-off you pay...

      Delete
    2. 'sampling' in a DVS sensor is something different than in an classical CIS. its about taking the decision that 'it has become brighter' or 'it has become darker' - so emit a bit and tag a timestamp to it. 5ns still seems to be a typo to me, at the moment the available DVS I know (but most likely my knowledge is a bit outdated...) are in the micorsecond range of 'sampling'.

      Delete
    3. let's do some calculation, assuming a f-1.4 lens, 10um pixel, 50% QE, pointing straight at the sun, how many photons can each pixel receive in 5ns? my calculation tells me it's in the range of 10s. and that's direct sunlight, what about office lighting, that's less than 1 photon per pixel in 5ns. so if my calculation is not way off, what they're claiming is not possible, regardless of what A-to-D conversion to use (DVS is basically a delta modulator array). how can you get any meaningful signal out of every fraction of 1 photon, or even 10s of photons (think about the shot noise)?

      Delete
    4. +1
      Agree with the above comment and calculations.

      Delete
    5. As we are talking about "event driven sensor" the frame rate has no meaning per se. Given enough pixels in the array, it is completely possible to have a pixel emitting an event every 5ns, even with an equivalent integration time of a few ms.
      But, once again, integration time has not the same meaning in this kind of sensor, as each pixel has its own integration time based on the "significant change in radiance" it has been programed with.

      Delete
    6. surely it is possible to have a pixel emitting an event every 5ns. but if given that at least a few micro-seconds of collected photons is needed to provide a meaningful SNR, then the events emitted at 5ns resolution are purely encoding shot noise, not "significant change in radiance".

      Delete
    7. Re "the events emitted at 5ns resolution are purely encoding shot noise": if the events are generated based on a equivalent integration time of say 1us or 1ms, then they can be meaningful signal. however, is it necessary or advantageous to have this 5ns time resolution when the equivalent integration time is 1000x larger? Probably not. Even though 1000x of Nyquist rate oversampling could lead to a higher SNR for the delta modulation, I doubt it is the right direction to push the DVS. I always see DVS as a different animal that trades SNR/details for speed and efficiency. For better SNR and details, we should not use delta modulation to begin with.

      Delete
  2. Interesting! The very first comment sounds as an University Professor (theoretical) who believes everything is possible; while as we go down, we hit ground reality!

    ReplyDelete
    Replies
    1. When tons of money is in front of you, everything is possible.

      Delete
    2. No; I have witnessed tons of money being spent with no desirable outcome.

      Delete
  3. The 5ns debate sounds like difference between accuracy and precision.

    ReplyDelete

All comments are moderated to avoid spam and personal attacks.