Hell Yeah! Direct Deepstack Integration - 5.4.0 - March 31, 2021

No configuration at all. Just took them out of the box, mounted them in the rack and started plugging in cables. Bought the first one through Amazon and the second from Nelly's. Excellent service from Nelly's by the way. My understanding of "auto negotiation" is that the devices look for the highest available speed at both ends and set the speed to that.

 
Oh dear, sebastiantombs and Edcfish, will I ever get to the bottom of this one? Firstly, please believe me when I say that it’s easy to understand that the data from all the camera ports in a switch must add arithmetically not to lose data in the upstream port. However, RevJoe’s advice to select a 1Gbit switch prompted me to sketch my system and ask the following question in post #426:-

I have one question - If I were to buy a (I'm guessing managed) switch specified at 1Gbit, do they all gather together data from their ports running at 100Mbit and upscale them into a 1Gbit stream if that upstream port is connected to a 1Gbit device (ie the network card)?

His answer was in post #427 was “NO”. He did introduce other considerations but my post #428 was an attempt to stick with the one key question. So what is the definitive answer? If RevJoe is correct and the 1Gbit upstream port in the switch auto-negotiates down to 100Mbit then we‘re back to my post #435.

Basic answer to your question is yes the data trying to get to your pc would be aggregated out that port, but in your case it can not all get out that port at the same time. So switches routes all the traffic it sees based on either IP/MAC address from all your cameras to one of your PC NICS based on your network settings and configurations. If you got 10 cams they all feed in to the switch then routed to that one port out that links your pc. In your case if all those cams are pushing 20MB each then your trying to push 200MB through straw that can only handle 100MB because the ports on your switch are all 100MB. If your Switch had a 1GB uplink port and your pc was connected to it your 200MB that's trying to get to your PC every second would flow without issue and you would be using 20% of your uplink speed. Your switch is the traffic cop telling the data it receives were to go based on the packet header info, but it can not keep the data moving because the uplink port/pipe to your pc is too small.

At a minimum I urge you to get switches that have a dedicated uplink of 1GB even if all the other ports (for your cams) are 100MB. Every switch has an internal network fabric/bandwidth that dictates how much data it can handle inside the switch, But when you have the majority of that data that comes into the switch going to only one outlet port that slows things down, especially if that port can not handle all the traffic trying to get out. Your work around will works but it is not efficient networking for a variety of reason. It does make sense to use switches in your out buildings but those switches should have a fast connections back to the mama ship. So at a minimum you should be buying switches that have a dedicate 1GB uplink port to use for your connection back to your server.

Hopefully we finally answered your questions. Networking is a very deep subject, but it always comes back down to bandwidth, throughput, and processing.
 
OK sebastiantombs, thanks for the reply. It would seem to me to be very do-able - you write the data from each camera port into a registers at 100Mbit and then read them sequentially at 1000Mbit for the upstream port. I wonder if some switches are clever enough to do that by detecting the 1Gbit NIC interface speed automatically without being ‘managed’. What about yours - did you have to configure it? Is there any chance, please that you could tell me the model number of your switches? Hopefully, there’s a manual on the Internet.
"Managed" just means the ability to actually configure individual switch ports. That means you can configure them for VLANS, STP, trunking, On/Off, etc... This is for advanced level networking. You don't need a "Managed" switch. You just need a basic (dumb) switch that has individual ports at Fast Ethernet speeds (10/100Mbps) and an uplink port at 1Gbps speed. Everything is auto negotiated. You don't have to do anything. If you plug that uplink into a router that supports 1Gbps, than you get 1Gbps speeds. If you plug it into a router that only supports 100Mbps, then you only get 100Mbps speeds. You don't have to configure or do anyting.
 
the switch does not need to be managed.
I have a BV-Tech POE switch that is unmanaged and works great.
It has 7 cameras and an WiFi Access Point attached to it's 100MB ports and the 1GB uplink connects to my 28 port DLink (which is then connected to my router/firewall).
 
Your currently static setting your NIC and cams correct? I get the impression you are, if so a unmanaged switch should work just fine.
 
I'm gearing up to do the same (I have i7-6700 16gb ram (adding more ram soon) - Curious if you're running BI and deepstack in Docker all on the same Windows 10 device or do you have your DeepStack in Docker running from a separate device?
All on the same machine. Deepstack can be run with the windows version or as a container in docker. I have the container with portainer also. Using the VorlonCD ai tool with this.
 
Dave Lonsdale, you also get AI switches today that offer a few DIP switches to enable certain modes. These are great because you don't need to log into them but they offer decent abilities like rebooting a PoE end device if no data transmission is detected or prioritizing video transmission. You don't need a managed switch per se.
 
  • Like
Reactions: Dave Lonsdale
Guess what guys, I realised I already have an old TP-LINK TL-SG1008D 8 port gigabit switch that an IT consultant gave me way back. It doesn’t have PoE so I hadn’t used it. To prove all gigabit switches auto negotiate the speed, I plugged my 100Mbit PoE switches into this one and hey presto, the link to the PC NIC runs at 1Gbit. Magic! Thanks again for your help.
 
Wow latest 5.4.6.0 now has the new option to test & tune.

1621362667113.png
 
Yeah very cool feature.
Like the fact now we can run the clips instead of just a screenshot. I would however love to see a Training module built in where we can train deep stack with custom modules :D
 
  • Like
Reactions: sebastiantombs
Like the fact now we can run the clips instead of just a screenshot. I would however love to see a Training module built in where we can train deep stack with custom modules :D

Yup! One thing I'm noticing is its scanning objects outside of the zone I created even when they aren't crossing into it. I thought it would scan after being triggered from crossing into the zone..
 
  • Like
Reactions: tech101
After migrating from AI tools, BI is missing many of the push notifications alters, but if I go back and analyze the clips with deepstack, it identifies a person. I have changed to 3 photos for 5 seconds in the settings, but it seems like it may not be long enough?


oof, I think it is because I capitalized the things I was searching for, ex. 'Person,Dog,Cat' and not 'person,dog,cat'.....
 

Attachments

  • Screen Shot 2021-05-20 at 10.07.55 PM.png
    Screen Shot 2021-05-20 at 10.07.55 PM.png
    28.7 KB · Views: 57
  • Screen Shot 2021-05-20 at 10.09.04 PM.png
    Screen Shot 2021-05-20 at 10.09.04 PM.png
    52.6 KB · Views: 57
  • Screen Shot 2021-05-20 at 10.10.43 PM.png
    Screen Shot 2021-05-20 at 10.10.43 PM.png
    41.7 KB · Views: 53
Last edited:
  • Like
Reactions: tech101
After migrating from AI tools, BI is missing many of the push notifications alters, but if I go back and analyze the clips with deepstack, it identifies a person. I have changed to 3 photos for 5 seconds in the settings, but it seems like it may not be long enough?


oof, I think it is because I capitalized the things I was searching for, ex. 'Person,Dog,Cat' and not 'person,dog,cat'.....
Yeah, as of 5.4.6.2, even faces that are capitalized in the system ("Michael" instead of "michael") don't trigger alerts. Seems like a bug, reported it.
 
  • Like
Reactions: tech101
Strangely I had a few cancelled alerts for objects that get correctly labeled when using test and tune. Wonder if it is just due to the particular images it sent to deepstack at the 1 second intervals not catching it. Wonder if it would be more accurate if we could lower the interval down to 0.5 second instead.
 
The problem with intervals shorter than one second is that a complete frame is being sent every second assuming the frame and iframe rates match. Shorter than one second would result in partial frames being analyzed.