Camera Selection - AI vs Deepstack

Looking Out

Pulling my weight
Joined
Feb 27, 2022
Messages
67
Reaction score
145
Location
New Jersey
I'm laying out my camera needs for a new system at my house with BI. Is there a real benefit to spending the extra money on cameras with AI versus non-AI versions?

Specifically I am looking at 4K models, mostly turret versions such as the following for example:
Amcrest IP8M-T2499EW VS.
Amcrest IP8M-T2669EW-AI
Will I get better accuracy?
Does it help with system resources? I'm using an i7-7700 for the BI server.
I am very interested in keeping images of cars coming and going in our dead end street as well as people on the sidewalk. Will this help?
 
As an Amazon Associate IPCamTalk earns from qualifying purchases.

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
25,173
Reaction score
49,060
Location
USA
If you are concerned about accuracy and need good night vision (hint we all do because we don't live in stadium lighting), those are bad choices. Bad night vision = bad AI detection.

They are 8MP on a 1/2.7" sensor. Those are budget cams for people thinking more MP is better. A 2MP on the same sensor will kick its butt all night long.

It is simple LOL do not chase MP - do not buy a 4MP camera that is anything smaller than a 1/1.8" sensor. Do not buy a 2MP camera that is anything smaller than a 1/2.8" sensor. Do not buy a 4K (8MP) camera on anything smaller than a 1/1.2" sensor. Unfortunately, most 4k (8MP) cams are on the same sensor as a 2MP and thus the 2MP will kick its butt all night long as the 4k will need 4 times the light than the 2MP... 4k will do very poor at night unless you have stadium quality lighting (well a lot of lighting LOL).

In addition, those are 2.8mm cameras which may or may not fit your needs. Unless these cameras are on the mailbox, a 2.8mm on your house will be useless for IDENTIFY purposes of a vehicle and people on the sidewalk (unless they are 10 feet away). It will let you know at best what time someone went by (maybe but those sensors are small with 4K), but that is about it.

You need to identify the areas you want to cover and pick a camera designed to cover that distance. In some instances, it may be a 2MP or 4MP that is the right camera. DO NOT CHASE MP!!!

Here are my general distance recommendations, but switch out the Dahua 5442 series camera to the equivalent 2MP on the 1/2.8" sensor or equivalent Hikvision works as well.
  • 5442 fixed lens 2.8mm or the 4K/X - anything within 10 feet of camera OR as an overview camera
  • 5442 ZE - varifocal - distances up to 40-50 feet (personally I wouldn't go past the 30 foot range but I like things closer)
  • 5442 Z4E - anything up to 80-100 feet (personally I wouldn't go past 60 feet but I like things closer)
  • 5241-Z12E - anything from 80 feet to almost 200 feet (personally I wouldn't go past 150 feet because I like things closer)
  • 5241-Z12E - for a license plate cam that you would angle up the street to get plates up to about 175 feet away, or up to 220 with additional IR.
  • 49225 PTZ - great auto-track PTZ and in conjunction with an NVR or Blue Iris and the cameras above that you can use as spotter cams to point the PTZ to the correct location to compliment the fixed cams.
You need to get the correct camera for the area trying to be covered. A wide angle 2.8mm to IDENTIFY someone 40 feet away is the wrong camera regardless of how good the camera is. A 2.8mm camera to IDENTIFY someone within 10 feet is a good choice OR it is an overview camera to see something happened but not be able to identify who.


Regarding camera AI versus Deepstack, it depends on your use case. Many of us use the camera AI and find it superior (at the moment anyway), but Deepstack is improving everyday. But if your use case is simply you want it to trigger for a car or person, it works great.

If your use case is to AI other objects or you want BI to be able to label it was a person or vehicle or truck or whatever, then you need Deepstack.

I mainly use the camera AI, but a few of my cams have a specific use case that I also use Deepstack.

With the use of substreams and the power of your computer, whether the camera does the AI or the computer won't make much of a deal.

 

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
25,173
Reaction score
49,060
Location
USA
This is a 4K camera on a larger/better sensor - can you IDENTIFY this person at the street? The Amcrest 4K would be even worse It would be A LOT darker out at the street and you may not even see a person. Most of us would not use this field of view and a 2.8mm to IDENTIFY someone at that distance of 50 feet.


1647114450527.png



But a 2MP camera OPTICALLY zoomed and set for the street to IDENTIFY results in this:


2am perp.png

Which are you trying to capture - the date/time someone walked or drove by or IDENTIFY? Depending on that answer, the Amcrest may or may not be the right camera.
 
Last edited:

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,696
Location
New Jersey
I throw in another two cents on this one.

If you use DeepStack it does not require a 4K image. In fact Blue Iris reduces the resolution of the snapshots sent to DeepStack to, roughly, 720P, say under 1MP. Having 4K or 2K cameras has no real effect other than increasing CPU load to downsize the images used for processing in AI.

Remember, sensor sizes are fractions, inversely proportional. On top of that as the number of pixels increases the light received by each pixel is reduced. A 1/2.7" 8MP will receive about 1/4 the amount of light than a 1/2.7" 2MP making the 8PM version basically blind at night.

1/3" = .333" Great for 720P
1/2.8" = .357" (think a .38 caliber bullet) Great for 2MP
1/1.8" = .555" (bigger than a .50 caliber bullet or ball) Great for 4MP
1/1.2" = .833" (bigger than a 20mm chain gun round) Great for 8MP
 
Top