very interesting. The "mobile edge compute' story continues to expand, and I has seen changes in microsoft, amazon, intel (especially) in the change to edge v.cloud. If you think back to when the cloud platforms launched everyone was focused on reoccurring rev. and the "lower maintenance, no equipment on site" BUT cloud is still expensive, and all the marketing associated with "As cloud adoption grows, the cost of cloud will decrease" This isnt the case. A major factor is the amount of data, and the that so many systems need GPU's to process. (again cloud GPU processing is $$$)
I didn’t vote because I believe there is the likely option of a mix, one with the camera doing the easy processing complete and pushing metadata to the cloud and another where more complex processing of the metadata, not the image is sent to the cloud. This would allow Vergara access to the information to refine the algorithms.
Since Vergara keeps communication to the camera available centrally it would be good to make the algorithms programmable for updates.
I am holding out for the next generation of cameras powered by Intel OpenVINO, that will be a game changer for IoT/Surveillance and the what the device can do will blow you away. Image a camera with AI that has its own broadband, goodbye traditional recording platforms.
#3, a few years ago, I thought that Intel / OpenVino / Myriad / Movidius chips were going to be big in video surveillance but I have not seen that in practice. From decisions from Avigilon, Bosch, and Verkada, to name a few bigger players, Amberalla has largely won. My understanding from manufacturer contacts is that this is primarily because Ambarella handles both (core video processing and AI) while Intel Myriad chips are just doing AI. If I am wrong about this, anyone is invited to correct me and explain why, thanks!
Am I drinking the Intel Valley Kool-Aid from the firehose? I know they sort of have blown it the first go around but I still think that once the lick this in the IoT space then it will happen.
I still think that once the lick this in the IoT space then it will happen.
Maybe, at some level one can make that case for NVIDIA. They are a huge company and a clear leader in GPUs but they are non factor for analytics processed on the edge.
The question for Intel and NVIDIA is do they really care about winning the edge camera analytics market. It may just be too small for them relative to the various opportunities they have in self-driving cars and various other applications.
I'm not sure that the poll accurately captures all the options.
IMO, analytics are moving to GPUs these days. The GPUs can be in a lot of different locations, depending on a number of factors, but the key element is that in order to do more in-depth object and scene analysis, GPUs are going to be "the thing" for the next generation of analytics.
As Mike Miller mentions, one issue here is heat/power (and also just the general size of an adequate GPU that fits in a camera).
I expect we will see some kind of Moore's-law kind of progress with GPUs, but probably not to the point that they become the size/power equivalent of a modern SoC, though SoCs will get noticeably better as a side effect.
So, my instinct is that we are going to see continued movement to the cloud for the "general" layer of analytics - storing video (though selectively), and most importantly managing logins to lots of sites, providing easy/secure remote access, etc. But the analytics will be done on-premise. Most likely not fully in the camera, so maybe on a separate server or processor box, or maybe on a recorder of some kind. The answer will probably depend on the site and the platform (manufacturer) they choose.
If Verkada gets serious about analytics, they're going to wind up needing some kind of on-site unit to handle the heavier lifting than what you can do in a camera, which is going to move them closer and closer to traditional architectures.
So, my instinct is that we are going to see continued movement to the cloud for the "general" layer of analytics - storing video (though selectively), and most importantly managing logins to lots of sites, providing easy/secure remote access, etc. But the analytics will be done on-premise. Most likely not fully in the camera, so maybe on a separate server or processor box, or maybe on a recorder of some kind. The answer will probably depend on the site and the platform (manufacturer) they choose.
The poll says 5 years into the future, this seems like a description of a typical high end install today :)
The GPUs we have today are pretty bad choices for inference and analytics as they contain a huge amount of unnecessary circuitry for 3d rendering and ray-tracing etc. that we just don't need - but you're paying for it anyways. As bad as it is, it is what we have available.
By running analytics on the edge you don't need to decode video prior to analysis. You also get to do the analytics at the lowest possible latency and at uncompressed quality. So what you'll need is a dedicated chip capable of running a CNN at a decent clip (tensorcore).
By having the processing cost of the analysis being borne by the camera, it means that the system is much more scalable. Instead of building a massively expensive super PC to handle 100 cameras, each camera brings its own muscle to the system. If the super PC dies, you're basically blind. With 100 little workers, it is less critical if one or two dies at the same time.
There will always be a need for off-line and after-the-fact analysis. Analytics may detect people, gender, colors and so on, but it can't know everything that you might want to look for. In those cases we don't really have much choice - get a beefy "analytics processor". This analysis can probably be accelerated by excluding large swaths of the database by looking at the metadata generated by the cameras themselves.
I don't see as much value in cloud recording; basically self hosting a reasonably sized vault is not that big of a deal for medium installations, probably a lot cheaper too. For micro-installs, cloud storage makes sense (upstream data is manageable).
Just because I think edge based analysis, on-site recording and cloud based infrastructure management is the optimal way of doing things doesn't mean that this is how things will pan out.
I guess all we need to do now, is wait 20 years, and maybe, just maybe ONVIF will have a 200 page document describing a standard for interop for these things.
I know. That is what I am addressing. It is barely 1 upgrade cycle in security.
By running analytics on the edge you don't need to decode video prior to analysis.
I have some passing familiarity with analytics at the edge. It's great for first-level stuff (this is a person, this is some kind of vehicles), but we're still a long way off (>1 upgrade cycle at least) before the average edge camera can say "This is a male with a red shirt, blue pants, moving in an unusual pattern" and do that not just for 1 or 2 objects, but for 20+.
For the next 5 years, I do not see analytics enabled cameras to catch up to what the industry is doing today with GPU/server-based systems. I do expect to see cameras get smarter overall, and I expect to see the GPU/server nodes get smaller and cheaper.
For the next 5 years, I do not see analytics enabled cameras to catch up to what the industry is doing today with GPU/server-based systems.
I agree with that - the physical limitations of the camera will always be inferior to what you can do with a dedicated PC. No question.
But if we don't see stronger edge analysis I don't think it will be because a monolithic solution with a GPU is superior (considering cost). It will be because there is little or no real value in such features. The customer will enjoy the dog and pony show presented by way of an old-school PC/GPU pair, but then order a bog standard IP-VCR with a collection of dumb cameras.
So I guess in some weird roundabout way I agree with you. I'm just bummed out as to the reason.
For the next 5 years, I do not see analytics enabled cameras to catch up to what the industry is doing today with GPU/server-based systems.
For sure, if what you mean which is the absolute 'best' / most advanced analytics.
On the other hand, it is already becoming apparent that the most common analytics are increasingly being run accurately on the edge (such as people / vehicle / face detection, etc.).
that the most common analytics are increasingly being run accurately on the edge
What is (mostly) being done at the edge today with good reliability is object detection eg: "this is a person". What is not being done, in most cases, is breaking the detection down any further (Male, red shirt, blue pants, has bag, has hat). High quality tracking at the edge is hit or miss, but at a minimum has been proven to be doable (though cross-camera tracking at the edge in a fully edge-based system starts to get messy, something that is potentially better done centrally).
If the premise is that analytics will mostly be used for basic alerting (person crossing tripwire, vehicle loitering), and for storage management (only record video when valid objects present instead of pixel motion), then edge analytics make a lot of sense.
However, I think the slow but measurable growth of analytics at the edge is driving user awareness and demand in a way that will push us more towards deeper usage of analytics. By this I mean, using analytics for search, doing re-identification search (find other appearances of same object) and then re-identification alarms (tell me the next time the person who looks like this comes into view of a camera). Particularly in the search use cases, just searching for Person or Vehicle gets boring quickly. Being able to search for a yellow SUV, or a female that took an unusual path through an area is potentially more valuable, and will likely require a level of computation that is not practical at the edge, or will require analyzing multiple FOVs, making it more logical to do this in some centralized manner.
I see GPU usage increasing, but in particular in the next 5 years I do not see the market segment getting big enough that we get highly specialized surveillance GPUs in the mainstream. So, as Morten points out, we're stuck using GPUs that come with some additional baggage. To me, this all points to not having todays level of GPU compute available in an SoC that is suitable for being used in a surveillance camera and driving what I would call full scale edge analytics. However, we can squeeze more out of those suboptimal centralized GPUs by pushing some of the basic inference tasks to the edge. Let the camera pick out the people and vehicles, and let the centralized GPU do the deeper analysis, longer-term analysis (eg: unusual pattern detection, trend analysis) and cross-camera computations (show me all the FOVs where this person appeared).
Lastly, as maybe an additional clarification, if some of my above expectations come true, then what we call "analytics" is likely to shift from basic object detection to deeper-level analysis, which is why I do not think "analytics" will be predominantly edge based, though on a total device count basis we would have more intelligent cameras than centralized GPUs, which would technically make cameras the most common place for analytics to run in 5 years if we don't expect the definition of analytics to advance much. As a counter argument to my own point, if we get 1 solid Moores Law round of GPU advancement, it might become more cost effective to add GPU/Server-based analytics to existing systems (keeping in mind 5 years is barely 1 upgrade cycle, lots of already-installed gear) making the server-based analytics the most common in 5 years, but maybe not in 10 years...
Lastly, as maybe an additional clarification, if some of my above expectations come true, then what we call "analytics" is likely to shift from basic object detection to deeper-level analysis, which is why I do not think "analytics" will be predominantly edge based
This could be. My gut feel is that this will mostly remain in the enterprise space, like how Briefcam has been used in the past decade. I am not saying it's not valuable, I do think it's valuable but given the cost primarily to enterprise users.
That's why I think the most common location of analytics going forward will be the edge with the cloud or server for enterprise / government / critical infrastructure users who want to do sophisticated tracking or analysis.
Well, if they used an NVR this would not be an issue. Pack plenty of processing in it and no problem. Worst case, much easier to upgrade a single NVR than a bunch of cameras. But then they could not make up some dumb reason to charge 6 times the price that the cameras are even worth, plus a subscription.
That might solve one part of the issue, and I’m not a Verkada fan. Let’s take this example.
You are the IT manager for a City. You have to update the City systems and that will be a change from IPv4 to IPv6 as well as new VPN’s and IP schemes.
There are 400 buildings, 6,000 cameras and about a thousand clients. One of the issues will be a requirement for facial recognition and LPR on multiple sites and only certain users via browser and mobile will have access to that and you have to log when they have accessed it and be able to run a report for authorities.
With Verkada, if it works as sold, this would be damn near seamless.
How would a bunch of NVR’s or even multiple VMS solutions work in this environment and what would it cost in upgrades, licensing and time?
Let me repeat.....I’m not a fan of Verkada’s current pay or brick offering.
Or you are the IT manager and you want to add in other cameras or connect an existing access control or VMS or add in some other analytics, etc and then you are screwed.
A great part of their sales model is that they make it easy to get in and terrible to get out or use other things.
We use this for a lot of our customers and like it very much. Lots of great features and benefits. This is a 3rd party app as Mobotix actually released MX-Cloud which is a an Eagle Eye product made specifically for Mobotix. We are currently testing MX-Cloud at this time and also like it. Our hope is that we can use this one day and not need the Eagle Eye bridge but the cameras SD card instead as the bridge.
They say their solution is unique because they use dual cameras. Many other systems use dual cameras.
Agree
Disagree
Informative
Unhelpful
Funny
Create New Topic
Subscribe to IPVM Research to read the full report.
Why do I need to subscribe?
The IPVM Research Service includes products tests and shootouts plus competitive and financial analysis, helping decision-makers better evaluate purchasing, partnering, developing, and/or competing against companies in physical security.
Comments (27)
Undisclosed Manufacturer #1
very interesting. The "mobile edge compute' story continues to expand, and I has seen changes in microsoft, amazon, intel (especially) in the change to edge v.cloud. If you think back to when the cloud platforms launched everyone was focused on reoccurring rev. and the "lower maintenance, no equipment on site" BUT cloud is still expensive, and all the marketing associated with "As cloud adoption grows, the cost of cloud will decrease" This isnt the case. A major factor is the amount of data, and the that so many systems need GPU's to process. (again cloud GPU processing is $$$)
So a hybrid edge compute system makes sense
Create New Topic
Undisclosed Manufacturer #2
I didn’t vote because I believe there is the likely option of a mix, one with the camera doing the easy processing complete and pushing metadata to the cloud and another where more complex processing of the metadata, not the image is sent to the cloud. This would allow Vergara access to the information to refine the algorithms.
Since Vergara keeps communication to the camera available centrally it would be good to make the algorithms programmable for updates.
IMHO
Create New Topic
Undisclosed End User #3
I am holding out for the next generation of cameras powered by Intel OpenVINO, that will be a game changer for IoT/Surveillance and the what the device can do will blow you away. Image a camera with AI that has its own broadband, goodbye traditional recording platforms.
Create New Topic
Brian Karas
11/06/20 04:28pm
I'm not sure that the poll accurately captures all the options.
IMO, analytics are moving to GPUs these days. The GPUs can be in a lot of different locations, depending on a number of factors, but the key element is that in order to do more in-depth object and scene analysis, GPUs are going to be "the thing" for the next generation of analytics.
As Mike Miller mentions, one issue here is heat/power (and also just the general size of an adequate GPU that fits in a camera).
I expect we will see some kind of Moore's-law kind of progress with GPUs, but probably not to the point that they become the size/power equivalent of a modern SoC, though SoCs will get noticeably better as a side effect.
So, my instinct is that we are going to see continued movement to the cloud for the "general" layer of analytics - storing video (though selectively), and most importantly managing logins to lots of sites, providing easy/secure remote access, etc. But the analytics will be done on-premise. Most likely not fully in the camera, so maybe on a separate server or processor box, or maybe on a recorder of some kind. The answer will probably depend on the site and the platform (manufacturer) they choose.
If Verkada gets serious about analytics, they're going to wind up needing some kind of on-site unit to handle the heavier lifting than what you can do in a camera, which is going to move them closer and closer to traditional architectures.
Create New Topic
Undisclosed Manufacturer #2
In general I agree with Morten. The ability for the camera to process video and turn it into usable metadata will be the first step for anything.
Create New Topic
Kris Tibbitts
Well, if they used an NVR this would not be an issue. Pack plenty of processing in it and no problem. Worst case, much easier to upgrade a single NVR than a bunch of cameras. But then they could not make up some dumb reason to charge 6 times the price that the cameras are even worth, plus a subscription.
Create New Topic
Undisclosed Integrator #4
Mobotix has been an edge camera device for years.
Create New Topic
Michael Miller
Looks like Mobotix is adding the cloud. MxMSP
Create New Topic
John Honovich
Related: Verkada has now announced LPR.
Create New Topic