Hell Yeah! Direct Deepstack Integration - 5.4.0 - March 31, 2021
- Thread starter fenderman
- Start date
CCTVCam
Known around here
- Sep 25, 2017
- 2,815
- 3,683
Thanks.
I think BI might have to seriously consider this as I understand the Tesla cards are going to be discontinued as well. I believe the AX which are Ampere based are going to take the place of all existing cards. The greta news i the perofrmance hike is massive. The bad news is the starting price jumps from around £110 currently for a P400 to around £699 from what info I can find, for a bottom of the range AX2000. Probably good value for a CAD user. Not so much for CCTV use.
This NVIDIA requirement isn't a BI issue it is a DeepStack issue.
Last edited:
Or consider using the google coral like Frigate does for $70 for the USB and a 10th of the power consumption.
CCTVCam
Known around here
- Sep 25, 2017
- 2,815
- 3,683
Thanks but that requires a 2 way connection through my firewall 24/7 and I see that as a vulnerability. In addition, the bandwidth usage is going to possibly make other activities such as gaming compromised as 24/7 BI will be sending pictures to the cloud for anaylsis and receiving the results back.
Google coral is completely local. It doesn't send anything to the cloud. The hardware is very impressive and a lot less power-hungry than GPUs with quicker inference speeds.
CCTVCam
Known around here
- Sep 25, 2017
- 2,815
- 3,683
Maybe you can provide some comparison to Quadro in performance and if better, talk to the BI devs about plug and play support for one of the modules.
As many can't code, I'm guessing unless it can be made plug and play, then it's still academic.
I am seeing interference speeds around 12ms, with 4 cameras ranging from 1080p to 4k using frigate.
sebastiantombs
Known around here
How many are triggered simultaneously? How many images are you analyzing? Four cameras is not much of a test, plus if it's Google it's behind the eight ball already to me. I don't trust Google further than I could throw them.
C
A coral is capable of 100fps. I am running each camera at 5fps as any more FPS does not improve the recognition. It's also built differently than how deep stack. Frigate is searching motion zones and only sends areas with the motion for processing. I will some times add a 5th camera but I haven't bothered to do get in my attic yet.
sebastiantombs
Known around here
OK, that's all well and good but in a normal surveillance situation you're looking at, typically, 10-20FPS and anything from ten cameras on up. I've currently got a dozen using DS and get excellent ID in under 100ms typically examining a total of 15 images per trigger.
I am recording at full camera FPS. Anywhere from 15-30 depending on the camera. I am running object recognition at 5fps as that is what the dev of frigate recommended as he did not see any improvement in object detection running faster FPS. The object detection is done in real-time with frigate. 1 coral would be able to support 5 cameras at 20fps if you wanted to run it at full speed or 20 at the recommended 5FPS. They make coral boards with 1 to 16 TPUs if you want to scale up.
sebastiantombs
Known around here
Thanks, I'll stick with an integrated package since I already have enough NVidia cards to handle a few hundred cameras with no problem. Zero investment, less equipment footprint and zero Google footprint.
That's fine, I am just making a suggestion for future improvement of the software not trying to change anyone's mind on their process. Seeing the interference speeds I am getting vs what I have seen in this thread and the complaints about pricing and power consumption I figured I would toss out a solution to improve on those two things. The coral board was designed by Google, other than that it is using the same models that deep stack takes advantage of and nothing is sent to google period but I guess if that is such a sticking point.
CCTVCam
Known around here
- Sep 25, 2017
- 2,815
- 3,683
Thnaks for the info. From what Sebastian is saying, it sounds as if it has some way to go on mathcing the quadros.
I understand that you can have more TPU's. However, I also read a warning with the 2 tpu board to watch power consumption through the m.2 slot as it can peak at 3 Amps just for the two tpus and so has the potential to fry power stages further down the chain if the MB can't take the transient high power draws. I've bought a gaming motherboard with a good power stage, so in itself, not too much of a worry. However, anything with the potential to burn something or that's going to run components near their rated spec has me worried for longetivity. Maybe when the techs more mature...
I believe this was the board I viewed: M.2 Accelerator with Dual Edge TPU | Coral
Current draw is mentioned in the pdf here in section 3.1 power consumption: M.2 Accelerator with Dual Edge TPU datasheet | Coral
I understand that you can have more TPU's. However, I also read a warning with the 2 tpu board to watch power consumption through the m.2 slot as it can peak at 3 Amps just for the two tpus and so has the potential to fry power stages further down the chain if the MB can't take the transient high power draws. I've bought a gaming motherboard with a good power stage, so in itself, not too much of a worry. However, anything with the potential to burn something or that's going to run components near their rated spec has me worried for longetivity. Maybe when the techs more mature...
I believe this was the board I viewed: M.2 Accelerator with Dual Edge TPU | Coral
Current draw is mentioned in the pdf here in section 3.1 power consumption: M.2 Accelerator with Dual Edge TPU datasheet | Coral
Last edited:
Just curios how does 5fps work if you set the trigger for let's say every 150ms for 5 seconds? Maybe I'm being dumb but doesn't the fps need to be higher than the interval rate? I've got my cameras set at 8-10fps and trigger 250 ms for example
Blue iris with dd
Thanks
What are the model architectures on the left column referring to?
Are they specific models you need to download and utilize like this?
github.com
Some of the interference times are astonishing.
Are they specific models you need to download and utilize like this?
tpu/models/official/efficientnet/edgetpu at master · tensorflow/tpu
Reference models and tools for Cloud TPUs. Contribute to tensorflow/tpu development by creating an account on GitHub.
github.com