There is research going on right now to build native camera systems that emit data (at the hardware level) that is not dense but rather sparse; information is transmitted as "blocks of pixels that changed between frame two and three" rather than "here is the state of all pixels at frame two, and here is the state of all pixels at frame three". In essence, doing a small amount of delta compression within the hardware itself.
This seems like a really natural fit for an application such as this one; and could potentially cut the amount of data needed to be paid attention to down by orders of magnitude.
Many of these ideas are similar to what is used to compress video down in the first place; it would be interesting to see a combination of video compression technology and "region of interest" ML pruning, where the ML models are applied only to interesting regions, as identified by the compression codec. The codec already has done the analysis to figure out which portions of the video have changed, which portions of the video are simply shifted versions of information from the last frame, etc.... You could significantly speed up ML analysis by making use of that kind of information.
A huge advantage of them is that you have practically instant latency on the output, as opposed to restricted frame rates with motion blur from traditional cameras. Have a look at [1] for a few demos. The spinning bucket path reconstruction and drone visual odometry are my favourites.
[0] http://rpg.ifi.uzh.ch/research_dvs.html [1] https://www.youtube.com/watch?v=F3OFzsaPtvI
This is already what hardware video codecs do well. If you’re talking about Chronocam-style differential sensors, they have their specialized use cases, but this is not one of them. Bandwidth from your sensor to your codec is already obscenely high, and you’re on a moving platform anyway.
As far as what video to actually store or do further processing on, ML ROI absolutely makes sense. This is basically what the Google Edge TPU is for: https://cloud.google.com/edge-tpu/
I've recently been doing some world simulation for games, and found that you quickly start pushing RAM limits if you want to represent an Earth-sized planet at square kilometer precision. Turns out the Earth is pretty big. Mars is more reasonable, with about a quarter of the surface area.
I don’t know if that was really true even back then, but today it’s completely doable and done all the time (think Google Chat or Facebook Chat).
My point is, times change, it’s entirely possible that in 20 years storing that much data will be completely possible.
In 20 years, we might have relatively faster links, but the improvement rate of those is much, much slower than storage density.
[1] https://www.wired.com/story/how-to-build-a-space-communicati...
[2] https://www.theverge.com/2013/2/1/3940898/darpa-gigapixel-dr...
If US gov doesn't do it, China will.
With compression I don't think the storage per se would be a problem for the IC.
And if you're looking for that particular missing airplane, you could just look at the timestamped photo and see it there.
My googling shows 148.94 trillion square meters, with 18.6 trillion
We'll being generous to say "good enough to spot a car" is 2 pixels with 1 meter pixel size.
To be moderately confident it's a car, I'd guesstimate we'd need quite a bit higher resolution.
With no compression and RGB 24 bit encoding, that's 3 bytes/pixel, or 446 TB of data per "frame" of our realtime recording. Of course that could be compressed down a TON, but regardless it's a staggeringly large quantity of data to work with.
> some machine learning it should be possible to skip most of oceans forests and deserts (unless some unusual object appears in that area).
As far as I can think through, you'd need to process the data in the oceans, forests, and deserts if you want to find unusual objects. We can't do a shortcut and say only 1/8 of the world is inhabited/developed in some way and that's all we're going to monitor. Perhaps you take 1/4 or 1/8 of the data from the vast amounts of uninteresting data and look for larger anomalies.
No matter how you look at it, there'd be a ridiculous amount of data and processing required to do object detection on that scale.