YouTube has removed videos of robots fighting, citing “animal cruelty”
technologyreview.com
technologyreview.com
Fact is indeed stranger than fiction..
The amount of man power to monitor, disk space, IT infrastructure, maintenance... yea. Whenever any city claims this or wants this, the reality sets in real fast as "Oh my shit, we can't do this". A real security camera, one that can run 24/7 in outdoor conditions, costs minimum $200 (Chinese ones) or you're replacing them every 1-2 years. I have had plenty of people who said I was wrong nad price gouging them. They later always had to buy the cameras I recommend before the 2 year mark hits. Dahua has decent cameras (still dog shit especially when it comes to view distance) starting at $200 each. But they do last a decent amount of time. 5 to 8 years. $500-$800 is the magic range (my opinion) for good quality Chinese cameras. But same goes for Bosch, Axis, etc. But the camera is your cheapest cost. The shear cost of all the support equipment, infrastructure, electricity, manpower... Especially with a city, it'd need such a high bitrate to have any type of resolution that matters with all that movement. You run into petabytes weekly at the minimum for a place like Beijing. Oh, I totally forgot about the video management system. Licenses are per camera on most of them. You have to pay like $50-$200 per camera for the software. Then all the servers to run the federated system. Don't get me wrong, every integrator wants a city (especially Beijing size) to say "Hey, we want 100% coverage in our city". That's retirement in one job. Granted a multi year job, depending on city. Still. Fun coupons. Lots of them.
Let me put some numbers down. 300 million cameras they say. Let's say weak, crappy 2mp camera, H.264, 12 frames a second, moderate scene activity (motion tracking to trigger recording). Save for 14 days (most cities first want 30, but request 14 after they have heart attacks from the price of storage). That's already ~100GB for those 14 days. Video and meta data take a lot of space. Multiply that by 300 million. That's ~30 billion GB. ~30mil TB. 30,000 PB. That's petabytes. Of pure no-profit storage. Just a cost sink. WD Purple drive 12TB drives (I like seagate, but the constant simultaneous read/write of a VMS, WD Purples are seriously the only real way to go) require a minimum of 2,500,000 drives. Even if they got a good deal at $250 for the bulk quantity (they're $370 consumer, and I use to get about 15%-20% off for bulk), that's $625mil in hard drives alone. Don't forget all the servers. So... many... servers... But also, redundancy, Raid5 or Raid6 depending on the firm doing it. So you need even more drives, by the ten-thousands. Oh, and the extra drives due to failure from that size of a deploy.
An average security camera of decent quality is around 5mp to 8mp. 14.44 TB to 26 TB of storage required under the same terms. Per camera. I say that's the bare minimum to even tell anything from up on a pole unless there's a decent zoom lens on the 2mp.
So when China says dumb shit like 300mil cameras in ONE city. Bullshit. Just plain bullshit. Because I promise you this. If true, I would be the one selling those damn cameras. Even one street in Beijing. I'd take ownership of that and retire. Gladly. Screw that, just the copper or glass rope alone. I'll take that piece of that pie.
TLDR: I did the math, China is full of shit about 300mil cameras in Beijing.
As an additional data point, according to their own media, China spent in 2018 "the equivalent of $20 billion USD" purchasing equipment for domestic surveillance, about half of total purchases in the global market. They've been making similar (or much larger) expenditures every year. So, that probably covers a wide part of the infrastructure.
> 300mil cameras in ONE city
I think they mean in the whole country.
That's still an unimaginably huge amount of streaming video to process for facial recognition, storage, etc., as you've outlined.
> I'll take that piece of that pie
I imagine, like in most other countries, this "defense" spending props up many well-connected companies, jobs, political and professional careers.
But you mention facial rec. So, again, I left 2 years ago. Most of the systems I knew tied to VMSs ended up always being separate boxes that handled the actual facial rec. I mean, some did inline. There's a thing called edge analytics (industry keyword) to cameras. Meaning the processor on the camera can do motion detection, facial, object track, etc, then pump that info out to a VMS that can handle that specific type of data (nightmare compatibility issues in this industry, ONVIF works until it doesn't). And, basic stuff works great. Motion and object. Yea. Facial... it "works". By works... eh, 70% success rate was the best I ever saw. Which means lots of false positives and lots of never caught. Enough to where staff tells me "This shit is useless". That's if the VMS can handle said data, because it wasn't part of ONVIF 2 when I was around. Oh and, it lags the hell out of the camera and recording if on the same server. Enough to where you have missing video. Which is bad. Real bad.
The way around, separate servers and software with real world ~90% success. Good enough to where filtering through the false positives by someone is pretty easy. This... SUPER EXPENSIVE. Not little expensive. Really bloody fucking expensive to deploy in hardware alone. Petabytes of data being crunched every minute. I'd probably get bjs from both Nvidia and Intel for selling that job. Plus, the bigger the database of bad faces, the longer the crunching takes. I'm trying to remember what my one customer needed... I think it was a dual E5-2695 v4 (maybe it was v3) along with whatever the high end Quadro at the time wise. I think it was the P4000. I'm looking at parts at that time frame, but I'm not 100%. Either way, it was a $10k+ box to just crunch faces (and was above spec since I built it) on about 20 high quality floor cameras (10mp+ wall mounted cameras that had clear face shots). That box was running at around 50%-60% capacity on average days and 80%-90% at peak hours. I think the memory was around 64gb and would max at 50 or so at peak times. Hardware alone would be killer for this china deploy, even if they have inhouse facial rec for "free". And the power draw... God have mercy on that powergrid.
But there are sooooooo, soooooo, soooo many problems with such a deployment. To act like it's a "one system" 300mil camera deploy. No. It's a pain in the ass to do a 1,000 camera deploy for a casino. A real pain in the ass. Different VMSs do these types of deployments differently. But generally what happens is this, you have some master servers that control slave servers. The master servers are what staff log into. "Even if a slave server goes down, the staff can see all other servers just fine." Notice the quotes. You see those quotes. HA! That's brochure talk. I've dealt with most of the big VMS players minus Milestone but ONSSI 5 was built on Milestone so... whatever. God... that was such a cluster fuck. 5.0 was such a mess... Anyways, if a server goes down, it can hiccup the entire federated system. 9 times out of 10, not a big deal. But you can notice it. But the more machines, means that 1 out of 10, the master server can seize up. Which then can crash the slave servers. That's not good. Because they all have to reboot AFTER the master reboots. Then if you have separate archiving servers, these need to sync back up depending on how the redundancy is setup. Lots of downtime.
Look, I can share a lot of stories. But to save time and boring stories of me being on the line with tech support where the lvl 2 and 3 guys go "Yea, I know, it's been an on going issue the past few years. Here's a workaround to make the customer happy. Just feed them xyz bullshit.", here's how to best explain a large scale camera deployment.
Think of a clock. Lots and lots of gears, moving parts. But instead of metal, the gears are made of glass. Some of the gears are thicker and some are really thin. Some are tempered, others you wonder how the hell they didn't shatter from looking at it too hard. At first, it can work great. Maybe a gear or two breaks and you fix them and it works on the initial deploy. But over time, the gears break or wear down in weird ways. Ways where the clock hands still move. It may not be the correct time, but the hands are still moving. Which is what counts at the end of the paycheck, cough, I mean day. So the system "still works". Kind of. Until the gears are so fucked from neglect, the entire thing should just be trashed and you buy another whole clock (not pieces) with glass gears. This is the security industry in a nutshell. Unless you get the expensive stuff from Axis, MOOG or anything from Israel. I'll throw Bosch in there too...eh. There are some companies in Israel that shoot their cameras to prove they're tough. In front of you.
So, yea. Such a wide deployment, I still call bullshit. Especially with maintenance and infrastructure costs. Keeping that system up and running, both electricity, camera replacement, cleaning the housing lens (seriously, smog fucks up those housing lenses hardcore), and we're not even talking about training staff to monitor the stuff yet. It's so expensive with extremely little value. Yes, they're surveillance crazy. But spending so much money on capturing how many "dangerous" people? Really? The guys at Antwerp that get robbed periodically spend less hunting down people who steal $100+ mil in diamonds per person.
And hell, maybe they did deploy this in the way they're talking about. An honest communist...bah. Probably why they're in so much debt. China has been debt heavy for many years now. A project like this can really cripple whoever implements it. It's that long term cost that no one thinks about and seriously hurts a lot of companies. Plus, cameras seriously don't last longer than 10 years. It's a 24/7 "cheap" device. So, they have to replace those cameras every 10 years, EASY. Max, 10 years. If it's a good temperate area. Any sustained lows below 40F or above 95F causes a lot of damage to these things if they're not designed for it (meaning more $$$). You might spend $20b once for parts, but you'll be spending $0.5b to $1b to maintain it, every year on average.
Oh plus, last story. Seattle deployed a city wide mesh network with cameras a number of years ago. It was shut down, I think a week (maybe 2 weeks) into it being turned on. Why? One of the city councilmen (or some other city official of some important status) was caught with a hooker on that camera system and the news spread through the police department, fast. Because. You know. Video. Now, we have separate bodies of power in the country that allow this to happen. The cops spreading the video I mean. Well, hookers too, but that's not the point I'm making. China doesn't really look favorable on the idea of separate bodies of power. But, a true 100% coverage... it's bad for chinese politicians too. Especially ones with wandering peckers and loyalties.
Lots of relevant points, like the (un)reliability of the cameras; difficulty of integrating into a single system and keeping the servers running; processing and storing the data; on-going energy and maintenance costs; dealing with false positives; and (un)reliability of the humans running the whole thing.
So I suppose "Skynet" has a few more decades to be even close to covering the whole country of China. And machine learning is still far from achieving "intelligence" in any meaningful sense of the term.
I can't help but think, though, that given a long enough timeline (say, in the next century) these technologies do seem to have a possibility of merging into a semisentient technopocalypse.