Deepstack on two computers - one stream - different results.

[evercleardoc]

So I am new here and have been toying with BI and just got deepstack to work. I have two machines - one is the one that I plan to run 24/7 and another one that I will use as my daily driver to VPN into the other machine.

The 24/7 is the BI machine - It is a laptop with Core i7 8550U with 8GB of ram
The other machine is a Ryzen 7 3700X with 64GB of ram.
Confidence 25%, taking 5 images and analyzing for 1 sec each on Medium setting.

Once deepstack was working I exported the settings from one machine and then improted them into a new clean install of BI.

When deepstack runs, both machines are viewing the exact same stream. One machine will flag different parts of video, some will be flagged multiple times on one machine and on the other machine it will completely miss it. I'm running 8 cameras all at 1080P and the two machines record the same footage 24/7 continuous it is just so interesting how deepstack is identifying or "not" identifying some objects/people on one machine vs the other under the exact same settings... Only difference is the CPU.

Has anyone else come in to something like this?

 

[wittaj]

Not surprising at all. DeepStack can be CPU intensive depending on how the rest of your system is set up and the type of processor. And the images won't queue up for too many images before it just stops analyzing and timing out. In fact, people have seen significant improvements moving Deepstack to a GPU version instead of the CPU version.

Plus a laptop is not a good choice for a 24/7 BI machine as they are not designed to be under load 24/7. The U means it is set up as power efficient, so as the laptop gets hot, it will slow the CPU down. Running 24/7 it will constantly be throttling the CPU and that will impact BI and Deepstack performance.

The key would be the make time that it takes Deepstack to determine if an object was found or not. Look into the BI log files and see what those times are in ms and report back.

And what is the CPU % for both computers?

 

[evercleardoc]

Not surprising at all. DeepStack can be CPU intensive depending on how the rest of your system is set up and the type of processor. And the images won't queue up for too many images before it just stops analyzing and timing out. In fact, people have seen significant improvements moving Deepstack to a GPU version instead of the CPU version.

Plus a laptop is not a good choice for a 24/7 BI machine as they are not designed to be under load 24/7. The U means it is set up as power efficient, so as the laptop gets hot, it will slow the CPU down. Running 24/7 it will constantly be throttling the CPU and that will impact BI and Deepstack performance.

The key would be the make time that it takes Deepstack to determine if an object was found or not. Look into the BI log files and see what those times are in ms and report back.

And what is the CPU % for both computers?

Thanks for the input. Very much appreciated.
Here are snips from the log file:

The Intel i7 8550U laptop's deepstack performance is as follows (CPU usage is 20%-99% from 1.8ghz-4ghz, ram is 90%):
0 8/7/2021 8:46:33.089 AM Entryway DeepStack: person:88% [844,236 1171,743] 5657ms
0 8/7/2021 8:46:33.090 AM Entryway DeepStack: person:70% [1087,727 1227,1008] 5657ms
0 8/7/2021 8:46:33.106 AM Entryway DeepStack: person:53% [1787,341 1893,543] 5657ms


The Desktop Ryzen 3700x performance is (CPU usage is 5% - 30% at any given time - overclocked to 4.2 ghz. Ram is 26-30%.):
0 8/7/2021 8:45:35.460 AM LivingRoom DeepStack: person:53% [231,186 435,498] 460ms
3 8/7/2021 8:45:37.994 AM Hallway MOTION
0 8/7/2021 8:45:48.700 AM Hallway DeepStack: Alert cancelled (nothing found)
3 8/7/2021 8:46:01.970 AM Hallway MOTION
0 8/7/2021 8:46:03.398 AM Hallway DeepStack: person:53% [357,3 631,185] 275ms
3 8/7/2021 8:46:10.910 AM Entryway MOTION
3 8/7/2021 8:46:18.026 AM Kenny MOTION
0 8/7/2021 8:46:26.729 AM Kenny DeepStack: Alert cancelled (nothing found)
3 8/7/2021 8:46:26.740 AM LivingRoom MOTION
0 8/7/2021 8:46:49.742 AM Entryway DeepStack: Alert cancelled (nothing found)
0 8/7/2021 8:46:51.058 AM LivingRoom DeepStack: Alert cancelled (occupied)
3 8/7/2021 8:46:54.373 AM Entryway MOTION
0 8/7/2021 8:46:57.897 AM Entryway DeepStack: person:42% [613,500 989,757] 302ms



For kicks I tried running the laptop BI to use the Ryzen desktop's deepstack server and I got the times on the laptop down to 250-500 ms for each ID. Interestingly, the desktop and the laptop still had different points of identification for the same clips but it was pretty close to each other... still not the same but close. I'm guessing because it was the same CPU. I then put it back to the local server and the times went back up to 3000-5000 ms... I would use an NVIDIA card but I already have a Radeon and no more slots left.

It is reassuring the laptop can ID similarly... just taking 5 x the time.

Just for fun I tried Sentry AI and am doing a little experiment.

Scenario 1: Laptop with Sentry AI compared to Ryzen desktop with deepstack - Both pushing to different iphones and counting/reviewing only flagged alerts.
Scenario 2: Laptop with deepstack compared to Ryzen with Sentry AI - Again, both pushing to different iphones.

I think this will help answer if deepstack on the laptop is worth anything.

Thanks for the input and feedback.

 

[wittaj]

Yeah 5,600ms is way too much and right on the verge of timing out. Everything will suffer on the that CPU running that. I am sure the laptop is either maxing CPU or coming close during that time. From looking around the DS forum, it appears people like to try to keep them under 200ms, but many are in the 40-75ms. But that is obviously dependent on your hardware.

At 5,000ms plus, I wouldn't count on that CPU to work consistently and identify with DeepStack if there is a lot of activity going on.

Blue Iris runs everything sequentially, which is why it is so important to have the system optimized doing every optimization in the wiki as well as having a computer capable of running it. That is why the desktop and the laptop still have different points of identification for the same clips.

Here is how I have noticed with BI that it is a sequence of events and not simultaneous actions (but probably something to do with the CPU of the computer as well). I have a camera for LPR that takes an snapshot. If I clone the camera, the snapshots are off a split second between the two cameras. One would think they would be identical snapshots being a clone, but the clone master takes the picture first and then the clone camera takes a snapshot and the two snapshots are different.

So if a lot of activity is going on during the day - windy, clouds, etc. and lots of cameras are in a sensing state and then several cameras have DS kick-in to do an analysis, that split second can be the difference between AI recognizing an object or not depending on the field of view.