Saturday, August 03, 2013

Microsoft & Rice University on Low Power Sensors

Microsoft Research and Rice University presentation "Energy-Proportional Image Sensing for Continuous Mobile Vision" by Robert LiKamWa, Bodhi Priyantha, Matthai Philipose, Victor Bahl, and Lin Zhong expands on the image sensor low power importance for mobile application, already presented at Taiwan Mobisys 2013 conference in June and won a Best Paper Award there.

A Youtube video talks about the low power ideas in this research:

And this Youtube video shows the Mobisys'13 presentation itself:


  1. Kwang-Bo (Austin) Cho, who made the first sub-milliwatt CMOS APS chip, wrote a Ph.D. dissertation on low power CMOS image sensor design, including a whole chapter on low power strategies. It would be nice if he got a mention by these newcomers to the low power game. (A 1.2 V Micropower CMOS Active Pixel Image Sensor, USC May 2001 along with several publications). Austin is currently a senior manager at Aptina and was at the 2013 IISW.

  2. Summary - Save power by reducing resolution, frame rate, and turning things off when they are not needed ... revolutionary ideas to be sure.

    1. Well if these ideas are so obvious, do it already. Why are these ideas not on every imager shipping out there, and when I go into binning or skipping I don't save any power? When I reduce the frame rate, power barely changes, if at all. Either these are not so obvious, or someone/everyone is being lazy. (Or everyone thinks that customers are not actually interested)

    2. I also think most, if not all, of these power savings "ideas" are obvious or have already been published or demonstrated. Generally features such as this are customer driven, so if you are Apple and you want these features, chances are you will get them. Adding features to sensors for which there is little demand, even when everyone knows how to do it, is not really going to happen. It is not a matter of being lazy - and is often a matter of cost, from NRE to test to die size. You can also pay for a custom sensor if these features are so compelling to your product.

      I think that when sensors that run continuously on batteries become high enough (potential) volume, you will see some of these power-saving features be implemented.

  3. Thanks for the coverage and for the comments. I understand and appreciate that low power savings has been approached by the image sensor community. However, most optimizations have been for particular quality settings. In fact, it seems that most image sensor systems optimize peak performance power consumption. That's not enough for us in the mobile systems community, since for different resolutions, we'd really like it to have different power consumption implications. We can't necessarily use multiple sensors with different qualities because a device's size profile may be limited. But some computer vision algorithms need high resolution frames, while many may suffice with very low resolution frames.

    Mobile systems and computer vision communities have thus far been treating the image sensor as a black box. Need an image? Okay, but it costs this much to turn on the image sensor. We wanted to finally ask the question (similar to Anonymous's above): "Why does it cost that much to capture a stream of images on the image sensor when we only need a low resolution image?" So we did that exploration. I think the bulk of the contribution of the work is not towards the image sensor community, but rather to the mobile systems community, revealing that we can in fact, with simple driver-level tweaks, create energy-proportional behavior, such that a 30 fps 1MP sensor (200 mW) can run at 5 fps, 0.3 MP and only consume 10 mW.

    I don't really expect image sensor companies to ship these kinds of scalable sensors if we don't have a product to put them in. If our push towards continuous mobile vision takes off, then maybe that'll change. But for now, our simple "systems-level" techniques with current image sensors will be a great starting point for us.

    Please, I welcome questions and dialogue. You may direct them to my email address (roblkw [at] my website domain). As Dr. Fossum points out, I am indeed a newcomer to the field. I would love to hear your perspective as experts in the image sensor community.

  4. Also note that while the sensors may support power saving modes, the software drivers many not properly use them. I've seen many drivers that are lucky to be functional, nevermind optimized for power savings. The problem is a combination of poor sensor documentation, tight schedules, and software engineers that aren't thinking about power savings and aren't experts in sensor operational tradeoffs.

  5. I take atmost 3 photos a day with my mobile camera, 2 video chats a week. For "real" photography, I grab my SLR.
    On the other hand, those power hungry apps need a battery doctor for sure...
    The real question is: do customers and or mobile industry require low-power capable cameras, or are we inventing a reason for them to save 30 min of battery time?

  6. Part of the drive for low power is heat management, not just longer battery life.


All comments are moderated to avoid spam.